← Backtests2026-10-18 00:17:05 UTC · Refresh snapshot

News → volatility

SPY / QQQ / IWM: looking for advance signals of movement beyond the price already charged for options, not just high volatility. We separately test warnings for rare losses from selling options. The archive and analysis are separate from the live news feed.

Qwen text analysis · 2025 · 115039 / 115039 (100.0%) · complete

Topic classified: 62963 · recaps/opinions: 52076 · main-stage errors: 0 · reject-audit errors: 0. Annotation progress, not forecast accuracy. 2026-09-12 20:43:19 UTC

H13 · unusual wording × option cost · annual checks below; not a validated trading signal

SPY / QQQ / IWM, noon–15:55 ET. Fixed2024 training;2025/2026 were previously accessed, not pristine holdouts. Historical point-in-time availability is not verified.

News-ready dates: 155 / 252; potential final training / OOF dates: 154 / 71. These are source counts, not forecasts or accuracy; prices and target classes may reduce them.

News-source count gates

At least60 common training dates and60 prior OOF dates are needed. The source uses observed-catalog metadata and timed-title quality gates, not all Yahoo news; missing data is not a calm-market label.

Quarter start	Potential training dates	Potential forecast dates
2024-04-01	40	42
2024-07-01	83	25
2024-10-01	107	47

SHA256 04554a081cd9a992940a761d7506c992afe599ffe122c91b9cb8e708bc5167da

2024 model frozen. These are training counts, not forecast accuracy. Final common training dates 133; minimum prior OOF dates per ETF/model 71; actual development fits 6. Required60 dates for each count and both target classes in every ETF.

2025 · predictive candidate not confirmed · 147 common dates

ΔBrier 0.00002; nominal95% CI [-0.00048, 0.00060]. Enhanced minus baseline; lower is better. Frozen2024 per-ETF prior benchmark Brier 0.15303.

Model	Brier	Precision / recall	Alerts: rows / dates	Coverage
Baseline	0.17557	4.76% / 1.20%	21 / 15	4.76%
+ lexical novelty	0.17558	4.76% / 1.20%	21 / 15	4.76%

Target-positive observations / distinct dates 83 / 33; eligible ETF-day rows 441. The target is future five-minute RV above an ATM Gaussian cost proxy, not option profit.

Separate economic mark check — not a profitability claim

Intrinsic value at15:55 minus paid ask premium and$4 costs; stress costs total$12. No option sale, exercise or capital/funding simulation. Missing time value and execution frictions remain. Only the primary complete-common-date cohort is used.

Enhanced secondary gate not passed. This cannot rescue the predictive test.

Model	Marks / dates / unknown	Mean stress mark / premium;95% CI	Paid premium sum; stress mark sum	Stress sum / premium sum
Baseline	21 / 15 / 0	-20.23%; [-46.59%, 7.04%]	$3493.00; $-504.50	-14.44%
Enhanced	21 / 15 / 0	-20.23%; [-46.59%, 7.04%]	$3493.00; $-504.50	-14.44%

Percentages are relative to paid option premiums, not annual return on capital. The sum of premiums repeatedly counts each selected opportunity; it is not required portfolio capital.

Base marks, worst marks and concentration

baseline: mean base / premium -14.79%, [-40.66%, 12.39%]; base sum $-336.50; worst base / stress $-185.00 / $-193.00; stress total without best3 dates (diagnostic only) $-1132.50; their share of positive date marks 94.94%.

enhanced: mean base / premium -14.79%, [-40.66%, 12.39%]; base sum $-336.50; worst base / stress $-185.00 / $-193.00; stress total without best3 dates (diagnostic only) $-1132.50; their share of positive date marks 94.94%.

All ETF/quarter controls and conditional intervals

Descriptive subgroups cannot replace the pooled annual criterion.

Group	Dates	ΔBrier; 95% CI	Enhanced precision / recall	Enhanced stress mean / premium
SPY	147	0.00013; [-0.00040, 0.00079]	0.00% / 0.00%	-69.78%
QQQ	147	0.00014; [-0.00048, 0.00081]	11.11% / 3.57%	-18.97%
IWM	147	-0.00023; [-0.00079, 0.00035]	0.00% / 0.00%	-4.97%
Q1	39	0.00038; [-0.00066, 0.00147]	0.00% / 0.00%	-2.07%
Q2	34	0.00050; [-0.00033, 0.00146]	n/a / 0.00%	n/a
Q3	32	-0.00131; [-0.00200, -0.00059]	6.25% / 10.00%	-18.21%
Q4	42	0.00030; [-0.00025, 0.00079]	0.00% / 0.00%	-32.85%

baseline: precision95% CI [0.00%, 13.64%] (1000/1000); recall95% CI [0.00%, 4.23%] (1000/1000). TP/FP/FN/TN 1/20/82/338. Log loss 0.89389.

enhanced: precision95% CI [0.00%, 13.64%] (1000/1000); recall95% CI [0.00%, 4.23%] (1000/1000). TP/FP/FN/TN 1/20/82/338. Log loss 0.90530.

Calibration: fixed10 bins

Model / bin	N	Mean forecast	Observed rate
baseline / 0.0–0.1	150	2.32%	20.00%
baseline / 0.1–0.2	102	15.43%	20.59%
baseline / 0.2–0.3	118	24.85%	18.64%
baseline / 0.3–0.4	52	33.97%	17.31%
baseline / 0.4–0.5	19	43.02%	5.26%
baseline / 0.5–0.6	0	n/a	n/a
baseline / 0.6–0.7	0	n/a	n/a
baseline / 0.7–0.8	0	n/a	n/a
baseline / 0.8–0.9	0	n/a	n/a
baseline / 0.9–1.0	0	n/a	n/a
enhanced / 0.0–0.1	151	2.33%	19.87%
enhanced / 0.1–0.2	104	15.61%	21.15%
enhanced / 0.2–0.3	117	24.91%	17.95%
enhanced / 0.3–0.4	51	34.14%	17.65%
enhanced / 0.4–0.5	18	43.09%	5.56%
enhanced / 0.5–0.6	0	n/a	n/a
enhanced / 0.6–0.7	0	n/a	n/a
enhanced / 0.7–0.8	0	n/a	n/a
enhanced / 0.8–0.9	0	n/a	n/a
enhanced / 0.9–1.0	0	n/a	n/a
constant / 0.0–0.1	0	n/a	n/a
constant / 0.1–0.2	441	17.79%	18.82%
constant / 0.2–0.3	0	n/a	n/a
constant / 0.3–0.4	0	n/a	n/a
constant / 0.4–0.5	0	n/a	n/a
constant / 0.5–0.6	0	n/a	n/a
constant / 0.6–0.7	0	n/a	n/a
constant / 0.7–0.8	0	n/a	n/a
constant / 0.8–0.9	0	n/a	n/a
constant / 0.9–1.0	0	n/a	n/a

Ledger / forecast rows 750 / 441. feature_reasons: complete=443, market=10, news=297. target_reasons: complete=740, entry=10.

SHA256 080386007bd29a84c30efce40ad43bfa31168dc2f27225abd230921eca8ec2ef

2026 · predictive candidate not confirmed · 169 common dates

ΔBrier -0.00002; nominal95% CI [-0.00038, 0.00029]. Enhanced minus baseline; lower is better. Frozen2024 per-ETF prior benchmark Brier 0.16786.

Model	Brier	Precision / recall	Alerts: rows / dates	Coverage
Baseline	0.18934	11.48% / 6.54%	61 / 38	12.03%
+ lexical novelty	0.18932	13.24% / 8.41%	68 / 42	13.41%

Target-positive observations / distinct dates 107 / 46; eligible ETF-day rows 507. The target is future five-minute RV above an ATM Gaussian cost proxy, not option profit.

Separate economic mark check — not a profitability claim

Enhanced secondary gate not passed. This cannot rescue the predictive test.

Model	Marks / dates / unknown	Mean stress mark / premium;95% CI	Paid premium sum; stress mark sum	Stress sum / premium sum
Baseline	61 / 38 / 0	-35.50%; [-53.96%, -15.09%]	$12133.00; $-4020.01	-33.13%
Enhanced	68 / 42 / 0	-29.24%; [-48.90%, -7.95%]	$13765.00; $-3662.02	-26.60%

Percentages are relative to paid option premiums, not annual return on capital. The sum of premiums repeatedly counts each selected opportunity; it is not required portfolio capital.

Base marks, worst marks and concentration

baseline: mean base / premium -30.21%, [-48.71%, -9.89%]; base sum $-3532.01; worst base / stress $-344.00 / $-352.00; stress total without best3 dates (diagnostic only) $-4746.51; their share of positive date marks 61.00%.

enhanced: mean base / premium -24.07%, [-43.48%, -2.95%]; base sum $-3118.02; worst base / stress $-344.00 / $-352.00; stress total without best3 dates (diagnostic only) $-4768.52; their share of positive date marks 62.69%.

All ETF/quarter controls and conditional intervals

Descriptive subgroups cannot replace the pooled annual criterion.

Group	Dates	ΔBrier; 95% CI	Enhanced precision / recall	Enhanced stress mean / premium
SPY	169	-0.00012; [-0.00049, 0.00021]	21.43% / 8.11%	-14.10%
QQQ	169	0.00009; [-0.00023, 0.00040]	15.00% / 8.57%	-15.09%
IWM	169	-0.00004; [-0.00047, 0.00032]	8.82% / 8.57%	-43.80%
Q1	58	0.00010; [-0.00020, 0.00038]	25.00% / 13.95%	-28.77%
Q2	62	0.00007; [-0.00046, 0.00052]	5.26% / 2.44%	-21.88%
Q3	49	-0.00028; [-0.00108, 0.00052]	8.00% / 8.70%	-35.29%
Q4	0	n/a

baseline: precision95% CI [1.56%, 25.01%] (1000/1000); recall95% CI [0.88%, 14.00%] (1000/1000). TP/FP/FN/TN 7/54/100/346. Log loss 0.62165.

enhanced: precision95% CI [3.28%, 26.57%] (1000/1000); recall95% CI [1.87%, 16.53%] (1000/1000). TP/FP/FN/TN 9/59/98/341. Log loss 0.62065.

Calibration: fixed10 bins

Model / bin	N	Mean forecast	Observed rate
baseline / 0.0–0.1	56	4.76%	17.86%
baseline / 0.1–0.2	106	15.84%	28.30%
baseline / 0.2–0.3	151	25.34%	19.87%
baseline / 0.3–0.4	139	34.63%	21.58%
baseline / 0.4–0.5	48	44.35%	14.58%
baseline / 0.5–0.6	5	52.41%	0.00%
baseline / 0.6–0.7	2	63.23%	0.00%
baseline / 0.7–0.8	0	n/a	n/a
baseline / 0.8–0.9	0	n/a	n/a
baseline / 0.9–1.0	0	n/a	n/a
enhanced / 0.0–0.1	54	4.62%	18.52%
enhanced / 0.1–0.2	109	15.78%	27.52%
enhanced / 0.2–0.3	148	25.37%	20.27%
enhanced / 0.3–0.4	142	34.64%	21.13%
enhanced / 0.4–0.5	46	44.29%	15.22%
enhanced / 0.5–0.6	6	51.94%	0.00%
enhanced / 0.6–0.7	2	63.31%	0.00%
enhanced / 0.7–0.8	0	n/a	n/a
enhanced / 0.8–0.9	0	n/a	n/a
enhanced / 0.9–1.0	0	n/a	n/a
constant / 0.0–0.1	0	n/a	n/a
constant / 0.1–0.2	507	17.79%	21.10%
constant / 0.2–0.3	0	n/a	n/a
constant / 0.3–0.4	0	n/a	n/a
constant / 0.4–0.5	0	n/a	n/a
constant / 0.5–0.6	0	n/a	n/a
constant / 0.6–0.7	0	n/a	n/a
constant / 0.7–0.8	0	n/a	n/a
constant / 0.8–0.9	0	n/a	n/a
constant / 0.9–1.0	0	n/a	n/a

Ledger / forecast rows 522 / 507. feature_reasons: complete=507, news=15. target_reasons: complete=522.

SHA256 af151c1a2f89c64aea68513b0a54fbf27c86aca9c9a7278ba8baab3109be0d7a

Fixed criteria and limitations

Predictive criterion: at least100 common dates,30 enhanced-alarm dates,30 positive ETF observations,950 defined precision AND recall bootstrap draws; upper95% paired ΔBrier bound below0. Economic criterion separately:30 observed selected dates, no unknown selected marks,950 defined draws and lower95% stress-mean bound above0.

Five-date circular blocks,1000 fixed-seed draws; ETFs stay together by date. Nominal intervals do not correct for all prior experiments. Lexical novelty is not semantic novelty or independent-event count. Neither criterion proves70–80% accuracy or a profitable option strategy.

H16 · do ordered headline phrases add a signal?

SPY / QQQ / IWM, noon NY. A fixed 23-input market, cost and news-count baseline versus the same model plus at most 64 learned word/phrase features. Vocabulary, weights and thresholds are frozen from 2024. No LLM; this is a wording association, not verification of a catalyst or its cause.

2025

Candidate not confirmed · 147 / 250 registered dates in the period · 441 ETF observations; 2025-01-02 — 2025-12-31.

Different positive dates / positive ETF observations: 33 / 83. Forecast gate: not passed; separate economic gate: not passed.

Target: future five-minute realized volatility through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV or a probability of option profit. Clipped ridge estimates are not inherently calibrated.

Model	Alerts / all	Precision	Recall	Brier	Log loss
Baseline	21 / 441	4.76% · [0.00%; 13.64%]	1.20% · [0.00%; 4.23%]	0.175569	0.893889
With phrases	28 / 441	14.29% · [0.00%; 38.10%]	4.82% · [0.00%; 13.34%]	0.195762	1.023187

Δ Brier versus baseline: 0.020193 · 95% [-0.001360; 0.043632]. Below zero is better.

Δ Brier versus frozen training prevalence: 0.042730 · 95% [0.016269; 0.068866]. Below zero is better.

Separate terminal marks and all-opportunity utility

Terminal intrinsic mark at 15:55, not an executed option exit: remaining time value, execution and assignment are not simulated. Base/stress costs are $4/$12 per pair. Returns use each entry’s paid premium, not account capital or an annual rate. Unknown selected marks remain unknown; skipped entries and fees are zero.

Policy	Known / unknown	Stress sum, USD	Selected mean	Mean across ALL opportunities
Baseline	21 / 0	$-504.50	-20.23% · [-46.59%; 7.04%]	-0.96% · [-2.39%; 0.32%]
With phrases	28 / 0	$-1728.00	-19.38% · [-43.99%; 1.80%]	-1.23% · [-2.81%; 0.09%]
Unfiltered	441 / 0	$-14212.39	-23.45% · [-36.12%; -8.73%]	-23.45% · [-36.12%; -8.73%]

Same-opportunity utility gain versus baseline: -0.27% · 95% [-1.77%; 1.15%].

Same-opportunity utility gain versus unfiltered: 22.22% · 95% [7.82%; 34.59%].

All THREE lower confidence bounds must exceed zero: selected stress mean, utility gain over baseline, and utility gain over unfiltered entries. A positive selected mean alone is not superiority.

baseline · stress sum without the best three dates (diagnostic only): $-1132.50.

enhanced · stress sum without the best three dates (diagnostic only): $-2522.00.

unfiltered · stress sum without the best three dates (diagnostic only): $-28706.85.

All ETFs and quarters; descriptive only

Quarter counts show complete common dates / registered calendar dates, not complete-month coverage. Missing dates are not quiet-market outcomes.

Group	Dates / calendar	Δ Brier / baseline	Δ Brier / constant
SPY	147	0.018765	0.039518
QQQ	147	0.018335	0.048679
IWM	147	0.023481	0.039993
Q1	39 / 60	0.007928	0.032950
Q2	34 / 62	-0.012955	0.014302
Q3	32 / 64	0.081847	0.118470
Q4	42 / 64	0.011442	0.017118

Calibration and preregistered evidence requirements

Forecast: 100 common dates, 30 distinct positive dates, 30 enhanced-alert dates, and 950 defined draws for enhanced precision/recall and both Brier differences. BOTH Brier upper bounds must be negative. Economics separately needs 100 dates, 30 selected-mark dates, no selected unknowns for any policy, 950 defined draws and all three positive lower bounds.

baseline

Probability bin	N	Forecast	Observed
0.00%–10.00%	150	2.32%	20.00%
10.00%–20.00%	102	15.43%	20.59%
20.00%–30.00%	118	24.85%	18.64%
30.00%–40.00%	52	33.97%	17.31%
40.00%–50.00%	19	43.02%	5.26%
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

enhanced

Probability bin	N	Forecast	Observed
0.00%–10.00%	164	1.93%	20.12%
10.00%–20.00%	70	15.05%	25.71%
20.00%–30.00%	80	24.83%	17.50%
30.00%–40.00%	60	35.28%	15.00%
40.00%–50.00%	41	43.52%	14.63%
50.00%–60.00%	18	53.77%	11.11%
60.00%–70.00%	2	60.76%	50.00%
70.00%–80.00%	2	78.51%	0.00%
80.00%–90.00%	3	86.86%	0.00%
90.00%–100.00%	1	91.15%	0.00%

constant

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	441	17.79%	18.82%
20.00%–30.00%	0	—	—
30.00%–40.00%	0	—	—
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Source coverage and independent audit

Full ledger / forecasts / common dates: 750 / 441 / 147.

Features: complete=443; market=10; news=297.

Targets: complete=740; entry=10.

SHA256 8acf4f177c4f90efe04637696b248227c7688450ff0d10625aeecf9771bdaeb6
Independent review SHA256 de55577d30309b411ac4086a9817dfc7111d7daf589d8260b2195b6ce8e6a093

2026

Candidate not confirmed · 169 / 174 registered dates in the period · 507 ETF observations; 2026-01-02 — 2026-09-09.

Different positive dates / positive ETF observations: 46 / 107. Forecast gate: not passed; separate economic gate: not passed.

Model	Alerts / all	Precision	Recall	Brier	Log loss
Baseline	61 / 507	11.48% · [1.56%; 25.01%]	6.54% · [0.88%; 14.00%]	0.189345	0.621647
With phrases	39 / 507	17.95% · [3.03%; 37.51%]	6.54% · [0.95%; 14.02%]	0.189979	0.769475

Δ Brier versus baseline: 0.000634 · 95% [-0.014680; 0.016369]. Below zero is better.

Δ Brier versus frozen training prevalence: 0.022116 · 95% [0.001665; 0.043878]. Below zero is better.

Separate terminal marks and all-opportunity utility

Policy	Known / unknown	Stress sum, USD	Selected mean	Mean across ALL opportunities
Baseline	61 / 0	$-4020.01	-35.50% · [-53.96%; -15.09%]	-4.27% · [-7.43%; -1.58%]
With phrases	39 / 0	$-2150.51	-28.63% · [-43.94%; -11.14%]	-2.20% · [-3.96%; -0.65%]
Unfiltered	507 / 0	$-24704.97	-22.82% · [-32.52%; -12.58%]	-22.82% · [-32.52%; -12.58%]

Same-opportunity utility gain versus baseline: 2.07% · 95% [-0.46%; 4.72%].

Same-opportunity utility gain versus unfiltered: 20.62% · 95% [11.04%; 29.81%].

All THREE lower confidence bounds must exceed zero: selected stress mean, utility gain over baseline, and utility gain over unfiltered entries. A positive selected mean alone is not superiority.

baseline · stress sum without the best three dates (diagnostic only): $-4746.51.

enhanced · stress sum without the best three dates (diagnostic only): $-2834.50.

unfiltered · stress sum without the best three dates (diagnostic only): $-29600.97.

All ETFs and quarters; descriptive only

Quarter counts show complete common dates / registered calendar dates, not complete-month coverage. Missing dates are not quiet-market outcomes.

Group	Dates / calendar	Δ Brier / baseline	Δ Brier / constant
SPY	169	0.006011	0.020047
QQQ	169	-0.002968	0.017663
IWM	169	-0.001141	0.028639
Q1	58 / 61	0.004034	0.014860
Q2	62 / 62	0.010325	0.031881
Q3	49 / 51	-0.015652	0.018349
Q4	0 / 0	—	—

Calibration and preregistered evidence requirements

baseline

Probability bin	N	Forecast	Observed
0.00%–10.00%	56	4.76%	17.86%
10.00%–20.00%	106	15.84%	28.30%
20.00%–30.00%	151	25.34%	19.87%
30.00%–40.00%	139	34.63%	21.58%
40.00%–50.00%	48	44.35%	14.58%
50.00%–60.00%	5	52.41%	0.00%
60.00%–70.00%	2	63.23%	0.00%
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

enhanced

Probability bin	N	Forecast	Observed
0.00%–10.00%	113	3.20%	15.93%
10.00%–20.00%	99	15.54%	22.22%
20.00%–30.00%	104	24.75%	18.27%
30.00%–40.00%	91	34.29%	26.37%
40.00%–50.00%	69	44.57%	27.54%
50.00%–60.00%	19	54.49%	15.79%
60.00%–70.00%	8	66.00%	12.50%
70.00%–80.00%	3	74.65%	33.33%
80.00%–90.00%	1	81.75%	0.00%
90.00%–100.00%	0	—	—

constant

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	507	17.79%	21.10%
20.00%–30.00%	0	—	—
30.00%–40.00%	0	—	—
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Source coverage and independent audit

Full ledger / forecasts / common dates: 522 / 507 / 169.

Features: complete=507; news=15.

Targets: complete=522.

SHA256 9ca3864d116db824e4497cda3b808e515c79cc7d308f731d13def324f079475e
Independent review SHA256 c1fd44648be7525cf1249ec2624b32ca8db2abf4bf6067f799a9465dade8bec1

Retrospective declared-time titles are not verified point-in-time news. Both years were previously exposed. Circular five-date bootstrap (1000 draws) keeps three ETFs together; intervals are nominal, not adjusted for the full research search. No result here establishes a profitable strategy or live 70–80% accuracy.

H17 · does recent retraining help?

SPY / QQQ / IWM, noon NY. Refit the same price baseline and phrase model on the latest 60 complete common dates within 120 scheduled sessions. The immediately preceding session is purged; half-days and gaps still count toward the 120-session boundary. Only past outcomes enter each fit. The fixed H16 model and its original results remain unchanged.

2025

Candidate not confirmed · 141 / 250 registered dates · 423 ETF observations; 2025-01-02 — 2025-12-31.

Distinct positive dates / positive ETF observations: 33 / 83. Forecast gate: not passed; separate economic gate: not passed.

Target: future five-minute realized volatility through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV or a probability of option profit.

Model	Alerts / all	Precision	Recall	Brier
Rolling baseline (23 inputs)	66 / 423	25.76% · [12.72%; 39.34%]	20.48% · [10.00%; 30.44%]	0.185488
Rolling text model	67 / 423	19.40% · [5.97%; 33.33%]	15.66% · [3.95%; 32.48%]	0.216422
Fixed 2024 text model	28 / 423	14.29% · [0.00%; 37.50%]	4.82% · [0.00%; 13.34%]	0.200785
Frequency over the same 60 dates	—	—	—	0.159792

Δ Brier · Rolling baseline (23 inputs): 0.030934 · 95% [0.010300; 0.053681].

Δ Brier · Fixed 2024 text model: 0.015636 · 95% [-0.024166; 0.057287].

Δ Brier · Frequency over the same 60 dates: 0.056630 · 95% [0.027132; 0.087308].

Each difference is rolling text minus its comparator; below zero is better. ALL THREE upper confidence bounds must be negative.

Separate terminal marks and four economic requirements

A 15:55 intrinsic mark, not an executed option exit or exercise simulation. Base/stress costs are $4/$12 per pair. Ratios use each entry’s paid premium, not account capital or annualized returns. Unknown selected marks remain unknown; skipped opportunities contribute zero without entry fees.

Policy	Known / unknown marks	Known stress sum, USD	Selected mean	Mean over ALL opportunities
Rolling baseline (23 inputs)	66 / 0	$-2250.40	-16.44% · [-41.89%; 12.30%]	-2.56% · [-6.34%; 2.22%]
Rolling text model	67 / 0	$-1738.00	-18.79% · [-52.13%; 13.62%]	-2.98% · [-7.14%; 2.34%]
Fixed 2024 text model	28 / 0	$-1728.00	-19.38% · [-40.35%; 2.47%]	-1.28% · [-2.93%; 0.14%]
Unfiltered	423 / 0	$-13013.89	-23.14% · [-35.29%; -6.84%]	-23.14% · [-35.29%; -6.84%]

Same-opportunity utility gain over Rolling baseline (23 inputs): -0.41% · 95% [-3.51%; 2.79%].

Same-opportunity utility gain over Fixed 2024 text model: -1.69% · 95% [-6.21%; 3.46%].

Same-opportunity utility gain over Unfiltered: 20.16% · 95% [8.07%; 31.50%].

ALL FOUR lower bounds must exceed zero: selected stress mean, and gains over the rolling baseline, fixed text model and unfiltered entries. A positive selected mean alone is not enough.

All ETFs and quarters; descriptive only

Common dates / registered calendar dates are shown, not complete-month coverage. Missing dates are not quiet-market outcomes; no group can rescue the annual criterion.

Group	Dates / calendar	Δ Brier / rolling baseline	Δ Brier / fixed text	Δ Brier / frequency
SPY	141	0.019964	0.013971	0.051025
QQQ	141	0.035362	0.010986	0.059028
IWM	141	0.037474	0.021952	0.059837
Q1	39 / 60	0.039528	-0.001798	0.042221
Q2	34 / 62	0.037268	0.099497	0.109542
Q3	32 / 64	0.033048	-0.067315	0.034902
Q4	36 / 64	0.013761	0.029056	0.041581

Calibration and fixed evidence requirements

Forecast: 100 common dates, 30 distinct positive dates, 30 rolling-text alert dates, and 950 defined draws for its precision/recall and all three Brier differences. Economics: 100 dates, 30 selected known-mark dates, no selected unknowns under any policy, and 950 draws for all four required bounds.

The six numeric 2024 alarm thresholds stay unchanged. They do not guarantee equal alert coverage or calibrated probabilities after retraining.

Rolling baseline (23 inputs)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	146	2.31%	17.81%
10.00%–20.00%	87	15.08%	24.14%
20.00%–30.00%	65	25.01%	16.92%
30.00%–40.00%	62	35.11%	12.90%
40.00%–50.00%	29	44.14%	27.59%
50.00%–60.00%	15	55.77%	20.00%
60.00%–70.00%	12	64.41%	25.00%
70.00%–80.00%	5	74.49%	40.00%
80.00%–90.00%	2	84.68%	50.00%
90.00%–100.00%	0	—	—

Rolling text model

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	188	1.60%	22.34%
10.00%–20.00%	38	14.24%	15.79%
20.00%–30.00%	69	25.26%	8.70%
30.00%–40.00%	38	35.15%	21.05%
40.00%–50.00%	29	45.95%	27.59%
50.00%–60.00%	22	54.72%	22.73%
60.00%–70.00%	21	64.52%	33.33%
70.00%–80.00%	9	75.36%	11.11%
80.00%–90.00%	6	82.19%	0.00%
90.00%–100.00%	3	99.22%	0.00%

Fixed 2024 text model

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	161	1.85%	20.50%
10.00%–20.00%	65	15.12%	27.69%
20.00%–30.00%	78	24.86%	17.95%
30.00%–40.00%	56	35.36%	16.07%
40.00%–50.00%	37	43.71%	16.22%
50.00%–60.00%	18	53.77%	11.11%
60.00%–70.00%	2	60.76%	50.00%
70.00%–80.00%	2	78.51%	0.00%
80.00%–90.00%	3	86.86%	0.00%
90.00%–100.00%	1	91.15%	0.00%

Frequency over the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	19	7.46%	15.79%
10.00%–20.00%	125	13.56%	15.20%
20.00%–30.00%	189	24.25%	21.16%
30.00%–40.00%	90	31.39%	23.33%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 750 / 423 / 141. Daily model fits: 282.

Input exclusions: complete=443; market=10; news=297.

Unknown outcomes: complete=740; entry=10.

Forecast statuses: current_features_unavailable=3; current_short_session=9; current_source_unavailable=297; insufficient_training_dates=18; ready=423.

SHA256 85c9f56b6ca79460dbfbae0f71b25c29eff26bb29cea2b85dc36f6818b98a204
Independent audit SHA256 7f0e16355dd54a31df6e5cbac2f89546360228b6073f39186a8d1c5053d0566e

2026

Candidate not confirmed · 169 / 174 registered dates · 507 ETF observations; 2026-01-02 — 2026-09-09.

Distinct positive dates / positive ETF observations: 46 / 107. Forecast gate: not passed; separate economic gate: not passed.

Target: future five-minute realized volatility through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV or a probability of option profit.

Model	Alerts / all	Precision	Recall	Brier
Rolling baseline (23 inputs)	76 / 507	31.58% · [15.15%; 48.75%]	22.43% · [10.52%; 35.35%]	0.188199
Rolling text model	59 / 507	23.73% · [10.00%; 40.75%]	13.08% · [5.55%; 21.93%]	0.204403
Fixed 2024 text model	39 / 507	17.95% · [3.03%; 37.51%]	6.54% · [0.95%; 14.02%]	0.189979
Frequency over the same 60 dates	—	—	—	0.173403

Δ Brier · Rolling baseline (23 inputs): 0.016204 · 95% [-0.005709; 0.034985].

Δ Brier · Fixed 2024 text model: 0.014424 · 95% [-0.010811; 0.040077].

Δ Brier · Frequency over the same 60 dates: 0.031000 · 95% [0.004862; 0.057107].

Each difference is rolling text minus its comparator; below zero is better. ALL THREE upper confidence bounds must be negative.

Separate terminal marks and four economic requirements

Policy	Known / unknown marks	Known stress sum, USD	Selected mean	Mean over ALL opportunities
Rolling baseline (23 inputs)	76 / 0	$-3731.13	-22.09% · [-41.68%; -0.61%]	-3.31% · [-6.40%; -0.10%]
Rolling text model	59 / 0	$-6261.88	-37.41% · [-55.08%; -14.30%]	-4.35% · [-7.25%; -1.51%]
Fixed 2024 text model	39 / 0	$-2150.51	-28.63% · [-43.94%; -11.14%]	-2.20% · [-3.96%; -0.65%]
Unfiltered	507 / 0	$-24704.97	-22.82% · [-32.52%; -12.58%]	-22.82% · [-32.52%; -12.58%]

Same-opportunity utility gain over Rolling baseline (23 inputs): -1.04% · 95% [-4.53%; 2.21%].

Same-opportunity utility gain over Fixed 2024 text model: -2.15% · 95% [-4.89%; 0.82%].

Same-opportunity utility gain over Unfiltered: 18.47% · 95% [8.79%; 27.48%].

ALL FOUR lower bounds must exceed zero: selected stress mean, and gains over the rolling baseline, fixed text model and unfiltered entries. A positive selected mean alone is not enough.

All ETFs and quarters; descriptive only

Common dates / registered calendar dates are shown, not complete-month coverage. Missing dates are not quiet-market outcomes; no group can rescue the annual criterion.

Group	Dates / calendar	Δ Brier / rolling baseline	Δ Brier / fixed text	Δ Brier / frequency
SPY	169	0.015325	0.018491	0.033497
QQQ	169	0.014309	0.016258	0.026110
IWM	169	0.018978	0.008524	0.033393
Q1	58 / 61	0.003187	0.007556	0.017795
Q2	62 / 62	0.033580	0.004555	0.033411
Q3	49 / 51	0.009626	0.035041	0.043580
Q4	0 / 0	—	—	—

Calibration and fixed evidence requirements

The six numeric 2024 alarm thresholds stay unchanged. They do not guarantee equal alert coverage or calibrated probabilities after retraining.

Rolling baseline (23 inputs)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	189	2.51%	19.58%
10.00%–20.00%	110	14.60%	20.91%
20.00%–30.00%	83	24.99%	15.66%
30.00%–40.00%	52	34.42%	21.15%
40.00%–50.00%	33	43.77%	18.18%
50.00%–60.00%	23	54.12%	39.13%
60.00%–70.00%	12	64.25%	50.00%
70.00%–80.00%	1	72.48%	100.00%
80.00%–90.00%	1	82.57%	100.00%
90.00%–100.00%	3	100.00%	0.00%

Rolling text model

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	176	1.41%	21.02%
10.00%–20.00%	69	14.57%	18.84%
20.00%–30.00%	87	25.03%	12.64%
30.00%–40.00%	75	34.88%	26.67%
40.00%–50.00%	46	45.10%	28.26%
50.00%–60.00%	35	55.19%	28.57%
60.00%–70.00%	7	65.92%	14.29%
70.00%–80.00%	5	73.30%	0.00%
80.00%–90.00%	2	83.68%	50.00%
90.00%–100.00%	5	96.26%	20.00%

Fixed 2024 text model

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	113	3.20%	15.93%
10.00%–20.00%	99	15.54%	22.22%
20.00%–30.00%	104	24.75%	18.27%
30.00%–40.00%	91	34.29%	26.37%
40.00%–50.00%	69	44.57%	27.54%
50.00%–60.00%	19	54.49%	15.79%
60.00%–70.00%	8	66.00%	12.50%
70.00%–80.00%	3	74.65%	33.33%
80.00%–90.00%	1	81.75%	0.00%
90.00%–100.00%	0	—	—

Frequency over the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	9	8.33%	22.22%
10.00%–20.00%	170	16.42%	27.06%
20.00%–30.00%	326	23.09%	17.79%
30.00%–40.00%	2	30.00%	50.00%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 522 / 507 / 169. Daily model fits: 338.

Input exclusions: complete=507; news=15.

Unknown outcomes: complete=522.

Forecast statuses: current_source_unavailable=15; ready=507.

SHA256 a5333659a43b0a176c8cd7c2d34c56bc1239298c7027026d701a841f9d920cea
Independent audit SHA256 94256f30a623b218ab55ac6b76e4e5d2059f55991c7a128fb2015b940c224ed8

Postdiagnostic study: both years were already examined. Historical declared-time titles are not verified point-in-time news. Five-date circular bootstrap (1000 draws) keeps ETFs together and conditions on saved predictions without retraining; overlapping training windows and prior research make intervals nominal, not pristine-holdout evidence. No live 70–80% accuracy or profitable strategy is established here.

H18 · joint state of three ETFs

SPY / QQQ / IWM, noon NY. One new 29-input ridge model: 23 own numeric inputs plus six fixed transforms of the other two ETFs. Training retains exactly the latest 60 eligible common dates within 120 scheduled sessions, with the preceding session purged. H17 baseline, text and frequency scores are copied unchanged; the original text-admitted cohort and all unavailable days remain unchanged.

The six additions summarize peers’ morning volatility, range, recent volatility, option premium and endpoint magnitude/disagreement. Shared averages also change ridge regularization: a gain would not establish unique peer causality, broad stock-market breadth or incremental news value.

2025

Candidate not confirmed · 141 / 250 registered dates · 423 ETF observations; 2025-01-02 — 2025-12-31.

Distinct positive dates / positive ETF observations: 33 / 83. Forecast gate: not passed; separate economic gate: not passed.

Protocol sample adequacy: forecast — sufficient; economics — sufficient.

Target: future five-minute realized volatility through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV or a probability of option profit.

Model	Alerts / all	Precision	Recall	Brier
H17 baseline (saved)	66 / 423	25.76% · [12.72%; 39.34%]	20.48% · [10.00%; 30.44%]	0.185488
Joint ETF state (23 + 6 inputs)	75 / 423	29.33% · [13.72%; 45.92%]	26.51% · [12.66%; 39.77%]	0.187554
H17 rolling text (saved)	67 / 423	19.40% · [5.97%; 33.33%]	15.66% · [3.95%; 32.48%]	0.216422
H17 frequency over the same 60 dates	—	—	—	0.159792

Δ Brier · H17 baseline (saved): 0.002066 · 95% [-0.008069; 0.013333].

Δ Brier · H17 rolling text (saved): -0.028868 · 95% [-0.053496; -0.005012].

Δ Brier · H17 frequency over the same 60 dates: 0.027762 · 95% [0.005759; 0.051162].

Each difference is the joint-state model minus its saved H17 comparator; below zero is better. ALL THREE upper confidence bounds must be negative.

Separate terminal marks and four economic requirements

Policy	Known / unknown marks	Known stress sum, USD	Selected mean	Mean over ALL opportunities
H17 baseline (saved)	66 / 0	$-2250.40	-16.44% · [-41.89%; 12.30%]	-2.56% · [-6.34%; 2.22%]
Joint ETF state (23 + 6 inputs)	75 / 0	$-663.94	-8.38% · [-39.91%; 24.02%]	-1.48% · [-6.50%; 4.63%]
H17 rolling text (saved)	67 / 0	$-1738.00	-18.79% · [-52.13%; 13.62%]	-2.98% · [-7.14%; 2.34%]
Unfiltered	423 / 0	$-13013.89	-23.14% · [-35.29%; -6.84%]	-23.14% · [-35.29%; -6.84%]

Same-opportunity utility gain over H17 baseline (saved): 1.08% · 95% [-1.84%; 5.51%].

Same-opportunity utility gain over H17 rolling text (saved): 1.49% · 95% [-2.59%; 6.03%].

Same-opportunity utility gain over Unfiltered: 21.65% · 95% [9.27%; 32.18%].

ALL FOUR lower bounds must exceed zero: selected stress mean, and gains over the saved H17 baseline, H17 text model and unfiltered entries. A positive selected mean alone is not enough.

All ETFs and quarters; descriptive only

Common dates / registered calendar dates are shown, not complete-month coverage. Missing dates are not quiet-market outcomes; no group can rescue the annual criterion.

Group	Dates / calendar	Δ Brier / H17 baseline	Δ Brier / H17 text	Δ Brier / frequency
SPY	141	0.003540	-0.016424	0.034601
QQQ	141	-0.003400	-0.038762	0.020265
IWM	141	0.006058	-0.031417	0.028420
Q1	39 / 60	0.004030	-0.035498	0.006723
Q2	34 / 62	0.006884	-0.030384	0.079158
Q3	32 / 64	-0.003120	-0.036168	-0.001266
Q4	36 / 64	-0.000002	-0.013763	0.027818

Calibration and fixed evidence requirements

Forecast: 100 common dates, 30 distinct positive dates, 30 joint-state alert dates, and 950 defined draws for its precision/recall and all three Brier differences. Economics: 100 dates, 30 selected known-mark dates, no selected unknowns under any of the four policies, and 950 draws for all four required bounds.

The candidate uses the original numeric baseline 80th-percentile cutoff, with strict greater-than comparison; all six historical thresholds stay unchanged. They do not guarantee equal alert coverage or calibrated probabilities after retraining.

H17 baseline (saved)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	146	2.31%	17.81%
10.00%–20.00%	87	15.08%	24.14%
20.00%–30.00%	65	25.01%	16.92%
30.00%–40.00%	62	35.11%	12.90%
40.00%–50.00%	29	44.14%	27.59%
50.00%–60.00%	15	55.77%	20.00%
60.00%–70.00%	12	64.41%	25.00%
70.00%–80.00%	5	74.49%	40.00%
80.00%–90.00%	2	84.68%	50.00%
90.00%–100.00%	0	—	—

Joint ETF state (23 + 6 inputs)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	136	2.29%	19.85%
10.00%–20.00%	92	15.27%	21.74%
20.00%–30.00%	77	25.39%	11.69%
30.00%–40.00%	47	35.01%	14.89%
40.00%–50.00%	25	44.56%	32.00%
50.00%–60.00%	25	54.73%	20.00%
60.00%–70.00%	12	63.60%	25.00%
70.00%–80.00%	2	74.03%	50.00%
80.00%–90.00%	5	83.70%	60.00%
90.00%–100.00%	2	94.16%	0.00%

H17 rolling text (saved)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	188	1.60%	22.34%
10.00%–20.00%	38	14.24%	15.79%
20.00%–30.00%	69	25.26%	8.70%
30.00%–40.00%	38	35.15%	21.05%
40.00%–50.00%	29	45.95%	27.59%
50.00%–60.00%	22	54.72%	22.73%
60.00%–70.00%	21	64.52%	33.33%
70.00%–80.00%	9	75.36%	11.11%
80.00%–90.00%	6	82.19%	0.00%
90.00%–100.00%	3	99.22%	0.00%

H17 frequency over the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	19	7.46%	15.79%
10.00%–20.00%	125	13.56%	15.20%
20.00%–30.00%	189	24.25%	21.16%
30.00%–40.00%	90	31.39%	23.33%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 750 / 423 / 141. Daily model fits: 141.

Input exclusions: complete=443; market=10; news=297.

Unknown outcomes: complete=740; entry=10.

Forecast statuses: current_features_unavailable=3; current_short_session=9; current_source_unavailable=297; insufficient_training_dates=18; ready=423.

SHA256 20bd6c9e3f9ed85762dae6235dda7a7d0200cc8fe8b0fe021ffa7541de59fe0c
Independent audit SHA256 a7bc45b4495ef46bd3302c615cf1cecb83c7db3bb0e902d64412d30b7c15bd65

2026

Candidate not confirmed · 169 / 174 registered dates · 507 ETF observations; 2026-01-02 — 2026-09-09.

Distinct positive dates / positive ETF observations: 46 / 107. Forecast gate: not passed; separate economic gate: not passed.

Protocol sample adequacy: forecast — sufficient; economics — sufficient.

Target: future five-minute realized volatility through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV or a probability of option profit.

Model	Alerts / all	Precision	Recall	Brier
H17 baseline (saved)	76 / 507	31.58% · [15.15%; 48.75%]	22.43% · [10.52%; 35.35%]	0.188199
Joint ETF state (23 + 6 inputs)	80 / 507	28.75% · [13.54%; 45.12%]	21.50% · [10.10%; 35.08%]	0.189055
H17 rolling text (saved)	59 / 507	23.73% · [10.00%; 40.75%]	13.08% · [5.55%; 21.93%]	0.204403
H17 frequency over the same 60 dates	—	—	—	0.173403

Δ Brier · H17 baseline (saved): 0.000856 · 95% [-0.005908; 0.007807].

Δ Brier · H17 rolling text (saved): -0.015348 · 95% [-0.035321; 0.004551].

Δ Brier · H17 frequency over the same 60 dates: 0.015652 · 95% [-0.009697; 0.039745].

Each difference is the joint-state model minus its saved H17 comparator; below zero is better. ALL THREE upper confidence bounds must be negative.

Separate terminal marks and four economic requirements

Policy	Known / unknown marks	Known stress sum, USD	Selected mean	Mean over ALL opportunities
H17 baseline (saved)	76 / 0	$-3731.13	-22.09% · [-41.68%; -0.61%]	-3.31% · [-6.40%; -0.10%]
Joint ETF state (23 + 6 inputs)	80 / 0	$-4969.13	-23.36% · [-43.93%; 0.06%]	-3.69% · [-7.36%; 0.01%]
H17 rolling text (saved)	59 / 0	$-6261.88	-37.41% · [-55.08%; -14.30%]	-4.35% · [-7.25%; -1.51%]
Unfiltered	507 / 0	$-24704.97	-22.82% · [-32.52%; -12.58%]	-22.82% · [-32.52%; -12.58%]

Same-opportunity utility gain over H17 baseline (saved): -0.37% · 95% [-2.51%; 1.88%].

Same-opportunity utility gain over H17 rolling text (saved): 0.67% · 95% [-2.89%; 4.25%].

Same-opportunity utility gain over Unfiltered: 19.14% · 95% [10.30%; 27.68%].

ALL FOUR lower bounds must exceed zero: selected stress mean, and gains over the saved H17 baseline, H17 text model and unfiltered entries. A positive selected mean alone is not enough.

All ETFs and quarters; descriptive only

Common dates / registered calendar dates are shown, not complete-month coverage. Missing dates are not quiet-market outcomes; no group can rescue the annual criterion.

Group	Dates / calendar	Δ Brier / H17 baseline	Δ Brier / H17 text	Δ Brier / frequency
SPY	169	-0.000098	-0.015423	0.018074
QQQ	169	0.002044	-0.012264	0.013846
IWM	169	0.000621	-0.018357	0.015036
Q1	58 / 61	-0.008295	-0.011482	0.006313
Q2	62 / 62	0.004983	-0.028597	0.004814
Q3	49 / 51	0.006465	-0.003161	0.040419
Q4	0 / 0	—	—	—

Calibration and fixed evidence requirements

H17 baseline (saved)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	189	2.51%	19.58%
10.00%–20.00%	110	14.60%	20.91%
20.00%–30.00%	83	24.99%	15.66%
30.00%–40.00%	52	34.42%	21.15%
40.00%–50.00%	33	43.77%	18.18%
50.00%–60.00%	23	54.12%	39.13%
60.00%–70.00%	12	64.25%	50.00%
70.00%–80.00%	1	72.48%	100.00%
80.00%–90.00%	1	82.57%	100.00%
90.00%–100.00%	3	100.00%	0.00%

Joint ETF state (23 + 6 inputs)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	219	2.61%	18.26%
10.00%–20.00%	100	15.27%	21.00%
20.00%–30.00%	53	24.19%	16.98%
30.00%–40.00%	57	34.60%	28.07%
40.00%–50.00%	39	44.52%	15.38%
50.00%–60.00%	15	54.40%	20.00%
60.00%–70.00%	17	62.56%	47.06%
70.00%–80.00%	3	73.99%	100.00%
80.00%–90.00%	1	84.07%	100.00%
90.00%–100.00%	3	100.00%	0.00%

H17 rolling text (saved)

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	176	1.41%	21.02%
10.00%–20.00%	69	14.57%	18.84%
20.00%–30.00%	87	25.03%	12.64%
30.00%–40.00%	75	34.88%	26.67%
40.00%–50.00%	46	45.10%	28.26%
50.00%–60.00%	35	55.19%	28.57%
60.00%–70.00%	7	65.92%	14.29%
70.00%–80.00%	5	73.30%	0.00%
80.00%–90.00%	2	83.68%	50.00%
90.00%–100.00%	5	96.26%	20.00%

H17 frequency over the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	9	8.33%	22.22%
10.00%–20.00%	170	16.42%	27.06%
20.00%–30.00%	326	23.09%	17.79%
30.00%–40.00%	2	30.00%	50.00%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 522 / 507 / 169. Daily model fits: 169.

Input exclusions: complete=507; news=15.

Unknown outcomes: complete=522.

Forecast statuses: current_source_unavailable=15; ready=507.

SHA256 83aba15634decc5636884c4e0a8069c980b179c726a665284bc51cc2ff2f89bf
Independent audit SHA256 29063024fc270f04ed5fedecfeda5952604ad43462cc0c7c67ae01fd0d980a3e

H19 · previous-session implied-volatility horizon slopes

SPY / QQQ / IWM, noon NY. Both freshly trained models use 23 own inputs plus the previous-session VIX level; the candidate adds log(VIX9D/VIX) and log(VIX/VIX3M). These are volatility indices across horizons, not a futures curve or an ETF option price.

NEW joint training cohort: the latest 60 dates with all ETF and macro inputs and known RV targets within 120 scheduled sessions; the preceding session is purged. Both models and the frequency control use exactly those dates. H17 current-day exclusions remain; training dates can differ from H17/H18, so their annual results are not a like-for-like comparison.

Source limitation: all three closes must come from the same immediately preceding exchange session. The local daily table is mutable and has no historical versions or publication timestamps. The audit checks captured bytes and permitted reads, not the provider’s historical availability. No gap filling, last-available substitution or source refresh between years; missing inputs can make the check insufficient.

2025

Candidate not confirmed · 141 / 250 calendar dates · 423 ETF observations; 2025-01-02 — 2025-12-31.

Distinct positive dates / positive ETF observations: 33 / 83. Predictive / economic sample adequacy: sufficient / sufficient.

Target: future five-minute RV from noon through 15:55 exceeds a Gaussian proxy for the paid ATM straddle plus $4 costs. This is not actual IV, executable gamma profit or the probability of option profit.

Model	Alerts / all	Precision	Recall	Brier
Fresh baseline · 24 inputs	84 / 423	25.00% · [10.66%; 40.00%]	25.30% · [11.27%; 39.75%]	0.181525
Horizon slopes · 26 inputs	77 / 423	23.38% · [10.94%; 36.67%]	21.69% · [10.66%; 32.05%]	0.181069
Frequency on the same 60 dates	—	—	—	0.159792

Δ Brier · Fresh baseline · 24 inputs: -0.000455 · 95% [-0.004830; 0.004005].

Δ Brier · Frequency on the same 60 dates: 0.021277 · 95% [0.001490; 0.043800].

Candidate minus control: lower is better. BOTH upper confidence bounds must be below zero.

Separate terminal marks: three economic requirements

15:55 intrinsic marks, not executed option exits or an exercise simulation. Base/stress costs are $4/$12 per pair. Percentages use each entry’s paid premium, not account capital or annualized returns. Selected unknown marks stay unknown; skipping contributes zero across the same full opportunity cohort.

Policy	Known / unknown	Known stress sum, USD	Selected mean	Mean over ALL opportunities
Fresh baseline · 24 inputs	84 / 0	$5682.59	-6.72% · [-43.37%; 46.59%]	-1.33% · [-7.82%; 10.13%]
Horizon slopes · 26 inputs	77 / 0	$-1723.50	-15.21% · [-37.38%; 10.29%]	-2.77% · [-6.27%; 1.89%]
Unfiltered	423 / 0	$-13013.89	-23.14% · [-35.29%; -6.84%]	-23.14% · [-35.29%; -6.84%]

Same-opportunity utility gain over Fresh baseline · 24 inputs: -1.43% · 95% [-8.62%; 2.84%].

Same-opportunity utility gain over Unfiltered: 20.37% · 95% [6.90%; 32.16%].

ALL THREE lower bounds must exceed zero: the selected stress mean and utility gains over the fresh baseline and unfiltered entries. A positive selected mean alone is insufficient.

All ETFs and quarters; descriptive only

Coverage is common dates / calendar dates, not complete-month coverage. No subgroup rescues the annual criterion.

Group	Dates / calendar	Δ Brier / baseline	Δ Brier / frequency
SPY	141	-0.000177	0.027058
QQQ	141	-0.002068	0.018455
IWM	141	0.000879	0.018319
Q1	39 / 60	-0.001827	0.000325
Q2	34 / 62	0.000788	0.060436
Q3	32 / 64	-0.004545	-0.005369
Q4	36 / 64	0.003491	0.030677

Calibration and fixed gates

Forecast: 100 common dates, 30 distinct positive dates, 30 candidate-alert dates, 950 defined draws for candidate precision/recall and both Brier differences. Economics: 100 dates, 30 selected known-mark dates, no selected unknowns in any of three policies, 950 draws for all three bounds.

Both scores use the same fixed original baseline 80th-percentile cutoff, strictly exceeded. This does not imply equal alert coverage or calibrated probabilities after daily retraining.

Fresh baseline · 24 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	150	2.47%	16.00%
10.00%–20.00%	71	15.61%	19.72%
20.00%–30.00%	74	24.51%	21.62%
30.00%–40.00%	48	35.45%	18.75%
40.00%–50.00%	43	44.23%	23.26%
50.00%–60.00%	20	55.19%	20.00%
60.00%–70.00%	8	63.05%	37.50%
70.00%–80.00%	5	74.84%	0.00%
80.00%–90.00%	2	85.83%	50.00%
90.00%–100.00%	2	92.57%	100.00%

Horizon slopes · 26 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	154	2.38%	15.58%
10.00%–20.00%	73	15.17%	16.44%
20.00%–30.00%	70	24.13%	25.71%
30.00%–40.00%	51	34.97%	21.57%
40.00%–50.00%	37	44.11%	21.62%
50.00%–60.00%	19	55.69%	21.05%
60.00%–70.00%	9	62.28%	33.33%
70.00%–80.00%	6	73.96%	16.67%
80.00%–90.00%	4	85.02%	50.00%
90.00%–100.00%	0	—	—

Frequency on the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	19	7.46%	15.79%
10.00%–20.00%	125	13.56%	15.20%
20.00%–30.00%	189	24.25%	21.16%
30.00%–40.00%	90	31.39%	23.33%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 750 / 423 / 141. Model fits (two per eligible day): 282.

Input exclusions: complete=443; market=10; news=297.

Unknown outcomes: complete=740; entry=10.

Forecast statuses; counts are ETF observations, not dates: current_features_unavailable=3; current_short_session=9; current_source_unavailable=297; insufficient_training_dates=18; ready=423.

SHA256 7a11e2ae2ed40dab3886a469dfce70694a5ae811d3a4e456fa13c210be368375
Independent audit SHA256 d8d11a260ed95bba1bdd1407f45bf05faeea378152cb55e846cfa2bb61d7edc4

2026

Candidate not confirmed · 137 / 174 calendar dates · 411 ETF observations; 2026-01-02 — 2026-09-09.

Distinct positive dates / positive ETF observations: 43 / 100. Predictive / economic sample adequacy: sufficient / sufficient.

Model	Alerts / all	Precision	Recall	Brier
Fresh baseline · 24 inputs	70 / 411	30.00% · [12.69%; 50.74%]	21.00% · [9.09%; 34.86%]	0.204430
Horizon slopes · 26 inputs	71 / 411	28.17% · [11.83%; 48.15%]	20.00% · [8.54%; 33.81%]	0.208318
Frequency on the same 60 dates	—	—	—	0.190683

Δ Brier · Fresh baseline · 24 inputs: 0.003888 · 95% [0.000194; 0.007620].

Δ Brier · Frequency on the same 60 dates: 0.017636 · 95% [-0.010622; 0.042296].

Candidate minus control: lower is better. BOTH upper confidence bounds must be below zero.

Separate terminal marks: three economic requirements

Policy	Known / unknown	Known stress sum, USD	Selected mean	Mean over ALL opportunities
Fresh baseline · 24 inputs	70 / 0	$-3078.39	-23.36% · [-52.17%; 11.19%]	-3.98% · [-9.42%; 1.73%]
Horizon slopes · 26 inputs	71 / 0	$-3034.39	-22.26% · [-50.88%; 13.19%]	-3.85% · [-9.48%; 2.07%]
Unfiltered	411 / 0	$-18707.13	-21.33% · [-32.62%; -9.37%]	-21.33% · [-32.62%; -9.37%]

Same-opportunity utility gain over Fresh baseline · 24 inputs: 0.13% · 95% [-0.63%; 0.95%].

Same-opportunity utility gain over Unfiltered: 17.49% · 95% [8.18%; 26.36%].

ALL THREE lower bounds must exceed zero: the selected stress mean and utility gains over the fresh baseline and unfiltered entries. A positive selected mean alone is insufficient.

All ETFs and quarters; descriptive only

Coverage is common dates / calendar dates, not complete-month coverage. No subgroup rescues the annual criterion.

Group	Dates / calendar	Δ Brier / baseline	Δ Brier / frequency
SPY	137	0.003907	0.019851
QQQ	137	0.003999	0.012999
IWM	137	0.003758	0.020056
Q1	58 / 61	0.003513	0.025798
Q2	62 / 62	0.002090	0.006319
Q3	17 / 51	0.011725	0.031059
Q4	0 / 0	—	—

Calibration and fixed gates

Both scores use the same fixed original baseline 80th-percentile cutoff, strictly exceeded. This does not imply equal alert coverage or calibrated probabilities after daily retraining.

Fresh baseline · 24 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	153	2.32%	20.92%
10.00%–20.00%	78	15.24%	26.92%
20.00%–30.00%	60	25.16%	23.33%
30.00%–40.00%	52	34.58%	25.00%
40.00%–50.00%	33	44.98%	15.15%
50.00%–60.00%	24	54.33%	20.83%
60.00%–70.00%	10	64.33%	90.00%
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	1	91.22%	100.00%

Horizon slopes · 26 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	148	2.19%	20.95%
10.00%–20.00%	80	14.99%	26.25%
20.00%–30.00%	62	25.37%	19.35%
30.00%–40.00%	50	34.74%	32.00%
40.00%–50.00%	34	44.85%	14.71%
50.00%–60.00%	25	54.65%	36.00%
60.00%–70.00%	10	63.58%	50.00%
70.00%–80.00%	1	71.84%	0.00%
80.00%–90.00%	1	89.73%	100.00%
90.00%–100.00%	0	—	—

Frequency on the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	9	8.33%	22.22%
10.00%–20.00%	166	16.38%	27.11%
20.00%–30.00%	233	22.83%	22.32%
30.00%–40.00%	3	30.00%	33.33%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Coverage, abstentions and independent audit

Full ledger / forecasts / common dates: 522 / 411 / 137. Model fits (two per eligible day): 274.

Input exclusions: complete=507; news=15.

Unknown outcomes: complete=522.

Forecast statuses; counts are ETF observations, not dates: current_macro_unavailable=96; current_source_unavailable=15; ready=411.

SHA256 4fdc73440552eafdf359ad979709f241991343043028ac36652baeaa6b060d46
Independent audit SHA256 d26b67e0051248e86efca06a7d46a89470a602030e23f22d5601d69d46f5baab

Postdiagnostic test: both years were previously examined; this is not a pristine holdout or verified point-in-time news/market history. Five-date circular bootstrap, 1000 draws, keeps ETFs together and conditions on saved predictions without refitting. Intervals are nominal, not adjusted for the preceding hypothesis search. No live accuracy or executable trading edge is established here.

H20 · opposite policy wording across headlines

SPY / QQQ / IWM, noon NY. Both fresh models include the original 23 inputs and four shares of tariff/Fed tightening or easing wording. The candidate adds one interaction: opposite directions within the same subject family, in different headlines. This tests that interaction beyond its four main effects, not all news value against a no-news model.

Lexical coexistence is NOT semantic contradiction, independent confirmation or evidence of the same actor, event or horizon. Reports, questions, expectations and conditional wording can match. Two directions in one headline contribute to neither directional share; negation suppresses the whole headline. All usable deduplicated headlines remain in the denominator.

Exactly the ORIGINAL H17 current-day cohort and latest 60 eligible training dates within 120 scheduled sessions; the immediately preceding session is purged, and half-days retain their calendar positions. Both 27/28-input models and the frequency control use the same 60 dates. No new match-count filter, replacement training dates or rescue of an original abstention.

Only the frozen current (10:00,12:00] NY title interval is used, under the declared publication/modification clock plus five minutes. Historical live availability and publisher independence are unverified. Missing news is unavailable, never a zero vector; zero word matches on a usable date are valid. Audits verify saved lineage and calculations, not semantic truth or historical provider availability.

2025

Candidate not confirmed · 141 / 250 calendar dates · 423 ETF observations; 2025-01-02 — 2025-12-31.

Distinct positive dates / positive ETF observations: 33 / 83. Predictive / economic sample adequacy: sufficient / sufficient.

Model	Alerts / all	Precision	Recall	Brier
Fresh baseline · 27 inputs	75 / 423	24.00% · [10.97%; 38.46%]	21.69% · [10.71%; 32.14%]	0.200864
Wording coexistence · 28 inputs	75 / 423	26.67% · [12.37%; 41.82%]	24.10% · [11.76%; 37.29%]	0.203228
Frequency on the same 60 dates	—	—	—	0.159792

Δ Brier · Fresh baseline · 27 inputs: 0.002364 · 95% [-0.003977; 0.009495].

Δ Brier · Frequency on the same 60 dates: 0.043436 · 95% [0.013029; 0.080014].

Candidate minus control: lower is better. BOTH upper bounds must be below zero.

Separate terminal marks: three economic requirements

15:55 intrinsic marks, not executed option exits, an exercise simulation or delta hedging. Base/stress costs are $4/$12 per pair. Percentages use each entry’s paid premium, not account capital or annualized returns. Selected unknown marks stay unknown; skipping contributes zero across the same full opportunity cohort.

Policy	Known / unknown	Known stress sum, USD	Selected mean	Mean over ALL opportunities
Fresh baseline · 27 inputs	75 / 0	$-4237.91	-23.95% · [-47.86%; 2.47%]	-4.25% · [-8.39%; 0.46%]
Wording coexistence · 28 inputs	75 / 0	$-812.85	-11.21% · [-37.19%; 13.97%]	-1.99% · [-5.81%; 2.94%]
Unfiltered	423 / 0	$-13013.89	-23.14% · [-35.29%; -6.84%]	-23.14% · [-35.29%; -6.84%]

Same-opportunity utility gain over Fresh baseline · 27 inputs: 2.26% · 95% [0.50%; 4.33%].

Same-opportunity utility gain over Unfiltered: 21.15% · 95% [8.72%; 32.25%].

ALL THREE lower bounds must exceed zero: the selected stress mean and utility gains over the fresh baseline and unfiltered entries. A positive selected mean alone is insufficient.

All ETFs and quarters; descriptive only

Coverage is common dates / calendar dates, not complete-month coverage. No subgroup rescues the annual criterion.

Group	Dates / calendar	Δ Brier / baseline	Δ Brier / frequency
SPY	141	-0.002655	0.043215
QQQ	141	-0.000046	0.042347
IWM	141	0.009793	0.044746
Q1	39 / 60	-0.001609	0.014655
Q2	34 / 62	0.011117	0.133384
Q3	32 / 64	0.000534	-0.006137
Q4	36 / 64	0.000027	0.033729

Calibration and fixed gates

Both scores strictly exceed the same fixed original baseline 80th-percentile cutoff. This does not imply equal alert coverage or calibrated probabilities after daily retraining.

Fresh baseline · 27 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	160	1.95%	17.50%
10.00%–20.00%	77	14.96%	18.18%
20.00%–30.00%	62	25.11%	25.81%
30.00%–40.00%	54	34.62%	14.81%
40.00%–50.00%	32	44.32%	31.25%
50.00%–60.00%	14	55.00%	7.14%
60.00%–70.00%	7	65.30%	28.57%
70.00%–80.00%	7	75.79%	28.57%
80.00%–90.00%	4	82.17%	50.00%
90.00%–100.00%	6	98.62%	0.00%

Wording coexistence · 28 inputs

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	159	1.85%	18.24%
10.00%–20.00%	81	14.99%	20.99%
20.00%–30.00%	57	24.97%	17.54%
30.00%–40.00%	56	35.15%	16.07%
40.00%–50.00%	29	44.46%	31.03%
50.00%–60.00%	14	54.38%	14.29%
60.00%–70.00%	8	64.67%	25.00%
70.00%–80.00%	6	74.52%	33.33%
80.00%–90.00%	3	82.75%	33.33%
90.00%–100.00%	10	98.22%	20.00%

Frequency on the same 60 dates

Score bin	N	Forecast mean	Observed rate
0.00%–10.00%	19	7.46%	15.79%
10.00%–20.00%	125	13.56%	15.20%
20.00%–30.00%	189	24.25%	21.16%
30.00%–40.00%	90	31.39%	23.33%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Lexical counts, once per date

All current headline representatives count once per date, not once per ETF or each reuse in training. Mixed and negation-suppressed headlines stay in the denominator. “Neither” means no qualifying pair for that family, not no real event. These diagnostics are not additional fitted features.

All forecast dates, including unknown future labels: 141 dates, 31040 headlines.

Positive / zero interaction dates: 7 / 134. Suppressed headlines / neither family exclusively directional: 1128 / 30854.

Family	Up only	Down only	Mixed	Neither	Suppressed
Tariffs / trade policy	47	32	0	29833	1128
Fed / rates	1	106	4	29801	1128

Common known-target dates: 141 dates, 31040 headlines.

Positive / zero interaction dates: 7 / 134. Suppressed headlines / neither family exclusively directional: 1128 / 30854.

Family	Up only	Down only	Mixed	Neither	Suppressed
Tariffs / trade policy	47	32	0	29833	1128
Fed / rates	1	106	4	29801	1128

Coverage, abstentions and independent audit

Full ETF ledger / forecasts / common dates: 750 / 423 / 141. Model fits (two per original-ready day): 282.

Input exclusions: complete=443; market=10; news=297.

Unknown outcomes: complete=740; entry=10.

Forecast statuses; counts are ETF observations, not dates: current_features_unavailable=3; current_short_session=9; current_source_unavailable=297; insufficient_training_dates=18; ready=423.

SHA256 e8b640e369ee7db8c72738ec656d9759d4347a31bc322b3b46b535142650cb6b
Independent audit SHA256 d08ce0fb84754d38712a8ee7364f78222e44e2374e5d6c2debca8f3f7f09c3af

2026

Not published: independently audited results are required.

Postdiagnostic test: both years were previously examined; this is not a pristine holdout or verified point-in-time news/market history. The 2026 check follows an independently audited 2025 result regardless of its sign. Five-date circular bootstrap, 1000 draws, keeps ETFs together and conditions on saved predictions without refitting. Intervals are nominal, not adjusted for the preceding hypothesis search. No live accuracy or executable trading edge is established here.

H15 · does unusual morning turnover add a signal?

SPY / QQQ / IWM. A fixed price-only model versus the same model with one extra input: 09:30–12:00 close×volume divided by the median of the previous 20 full sessions. This is a turnover proxy, not buying pressure or exact traded notional. Both models were trained on 2022–2024, then frozen.

2025

Observed common-date range: 2025-01-02 — 2025-12-31.

Predictive advantage not confirmed · 246 common dates / 738 ETF observations. 62 distinct positive-outcome dates.

Target: future five-minute realized volatility until 15:55 exceeds the Gaussian proxy implied by the paid ATM pair plus $4 costs. This is not an option-profit label.

Model	Selected / all	Precision	Recall	Brier
Prices only	217 / 738	23.04% · [13.63%; 32.75%]	33.33% · [19.86%; 45.29%]	0.167348
With volume	207 / 738	23.67% · [14.04%; 33.47%]	32.67% · [19.26%; 44.66%]	0.167063

Δ Brier versus price-only model: -0.000285 · 95% [-0.000547; -0.000042]. Below zero is better.

Δ Brier versus frozen base frequency: 0.001572 · 95% [-0.005899; 0.009671]. Below zero is better.

Separate hypothetical economics and comparisons

These are 15:55 intrinsic marks, not executed option sales. Residual time value is omitted. Stress costs=$12 per pair; base costs=$4. Percentages divide by option premium, not account capital and not annual return.

Selection	Known / unknown marks	Stress mark sum	Mean per selected entry	Mean per ALL opportunities
Prices only	217 / 0	$404.80	-13.41% · [-32.47%; 11.33%]	-3.94%
With volume	207 / 0	$1015.09	-12.13% · [-32.82%; 12.95%]	-3.40%
No selection	738 / 0	$-29257.59	-22.35% · [-30.74%; -13.30%]	-22.35%

Change in all-opportunity mean versus «Prices only»: 0.54% · 95% [0.02%; 1.22%].

Change in all-opportunity mean versus «No selection»: 18.95% · 95% [11.86%; 25.56%].

Skipped opportunities contribute zero, including zero fees; unknown selected marks never become zero. Positive selected marks alone do not prove superiority to either comparator.

Positive-mark criterion not passed.

Every ETF and quarter; descriptive only

Group	Dates	Δ Brier / prices	Δ Brier / base
SPY	246	-0.000277	0.000765
QQQ	246	-0.000248	0.004590
IWM	246	-0.000329	-0.000639
Q1	59	-0.000623	-0.016607
Q2	62	-0.000375	0.012960
Q3	63	-0.000099	0.006028
Q4	62	-0.000062	0.002956

Calibration, sample requirements and concentration

At least 100 common dates, 30 distinct positive dates, 30 volume-model alert dates and 950 defined precision AND recall bootstrap draws. Both Brier-comparison upper bounds must be below zero. Economics has a separate gate.

Prices only

Probability bin	N	Forecast	Observed
0.00%–10.00%	150	5.18%	18.00%
10.00%–20.00%	402	15.12%	19.90%
20.00%–30.00%	153	23.75%	20.92%
30.00%–40.00%	30	32.30%	36.67%
40.00%–50.00%	3	48.70%	0.00%
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

With volume

Probability bin	N	Forecast	Observed
0.00%–10.00%	146	5.13%	18.49%
10.00%–20.00%	402	15.08%	19.90%
20.00%–30.00%	156	23.67%	19.87%
30.00%–40.00%	31	32.25%	38.71%
40.00%–50.00%	3	48.76%	0.00%
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Frozen base frequency

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	738	14.64%	20.33%
20.00%–30.00%	0	—	—
30.00%–40.00%	0	—	—
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Prices only · stress sum without the three best dates (diagnostic, never a filter): $-14089.66.

With volume · stress sum without the three best dates (diagnostic, never a filter): $-13479.37.

No selection · stress sum without the three best dates (diagnostic, never a filter): $-43752.05.

Source coverage and audit

Ledger / price-ready / volume-ready / forecasts: 750 / 740 / 740 / 738.

Extra exclusions due to volume history: 0; unknown future targets among forecasts: 0.

SHA256 f819c6613c9fc6bfb3d6b6b55159e708c6db9a43ee54236285bf062fe9af72ce

2026

Observed common-date range: 2026-01-02 — 2026-02-20. Insufficient sample for the registered evidence criterion.

Original integer-volume version: fractional custom-bar volumes from 23 February fail its admission rule. Price bars exist; these are not zero-volume or missing-market days. The whole year cannot be judged from the retained subset. A fractional-volume extension is a separate experiment and does not replace this result.

Predictive advantage not confirmed · 34 common dates / 102 ETF observations. 14 distinct positive-outcome dates.

Target: future five-minute realized volatility until 15:55 exceeds the Gaussian proxy implied by the paid ATM pair plus $4 costs. This is not an option-profit label.

Model	Selected / all	Precision	Recall	Brier
Prices only	61 / 102	27.87% · [14.29%; 41.19%]	58.62% · [38.10%; 75.69%]	0.219338
With volume	57 / 102	28.07% · [14.58%; 40.98%]	55.17% · [33.33%; 72.97%]	0.219508

Δ Brier versus price-only model: 0.000170 · 95% [-0.000170; 0.000557]. Below zero is better.

Δ Brier versus frozen base frequency: -0.002276 · 95% [-0.021040; 0.012753]. Below zero is better.

Separate hypothetical economics and comparisons

Selection	Known / unknown marks	Stress mark sum	Mean per selected entry	Mean per ALL opportunities
Prices only	61 / 0	$-2840.08	-20.42% · [-51.76%; 18.58%]	-12.21%
With volume	57 / 0	$-2872.57	-20.67% · [-53.21%; 18.30%]	-11.55%
No selection	102 / 0	$-4555.90	-19.09% · [-44.09%; 9.67%]	-19.09%

Change in all-opportunity mean versus «Prices only»: 0.66% · 95% [-1.76%; 3.08%].

Change in all-opportunity mean versus «No selection»: 7.53% · 95% [-11.71%; 25.36%].

Skipped opportunities contribute zero, including zero fees; unknown selected marks never become zero. Positive selected marks alone do not prove superiority to either comparator.

Positive-mark criterion not passed.

Every ETF and quarter; descriptive only

Group	Dates	Δ Brier / prices	Δ Brier / base
SPY	34	0.000361	0.001521
QQQ	34	0.000163	0.001568
IWM	34	-0.000014	-0.009918
Q1	34	0.000170	-0.002276
Q2	0	—	—
Q3	0	—	—
Q4	0	—	—

Calibration, sample requirements and concentration

Prices only

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	49	16.32%	34.69%
20.00%–30.00%	40	23.95%	25.00%
30.00%–40.00%	13	33.15%	15.38%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

With volume

Probability bin	N	Forecast	Observed
0.00%–10.00%	1	9.78%	0.00%
10.00%–20.00%	51	16.55%	33.33%
20.00%–30.00%	39	24.42%	25.64%
30.00%–40.00%	11	33.60%	18.18%
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Frozen base frequency

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	102	14.64%	28.43%
20.00%–30.00%	0	—	—
30.00%–40.00%	0	—	—
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Prices only · stress sum without the three best dates (diagnostic, never a filter): $-4723.07.

With volume · stress sum without the three best dates (diagnostic, never a filter): $-4695.07.

No selection · stress sum without the three best dates (diagnostic, never a filter): $-6765.38.

Source coverage and audit

Ledger / price-ready / volume-ready / forecasts: 522 / 522 / 102 / 102.

Extra exclusions due to volume history: 420; unknown future targets among forecasts: 0.

SHA256 2e7e12090955ee0a270428d3cdd2def18dd6e04894ab756998d818ece193d1be

1000 circular five-date bootstrap draws keep the three ETFs together. Intervals are nominal and do not adjust for all hypotheses tried. Previously studied years and corrected historical bars are not a pristine holdout or proof of live 70–80% accuracy. Prospective confirmation is still needed.

H15-FV · fractional volumes, unchanged models

Separate source-measurement extension: exactly represent volumes in millionths of a share, without rounding, new training or threshold selection. All original 2022–2025 integer-volume feature vectors and saved scores were reproduced bit for bit. The original H15 results remain visible and unchanged. This 2026 check follows inspection of the volume-reporting change: it is not a pristine holdout. Accepting fractional values does not prove that old and new volume measurements are economically identical; the reporting-regime shift remains unresolved.

2026

Observed common-date range: 2026-01-02 — 2026-09-11.

Predictive advantage not confirmed · 174 common dates / 522 ETF observations. 47 distinct positive-outcome dates.

Target: future five-minute realized volatility until 15:55 exceeds the Gaussian proxy implied by the paid ATM pair plus $4 costs. This is not an option-profit label.

Model	Selected / all	Precision	Recall	Brier
Prices only	245 / 522	25.71% · [17.67%; 34.25%]	57.27% · [44.66%; 68.18%]	0.163450
With volume	228 / 522	25.88% · [17.49%; 34.77%]	53.64% · [40.65%; 65.08%]	0.163589

Δ Brier versus price-only model: 0.000139 · 95% [-0.000089; 0.000380]. Below zero is better.

Δ Brier versus frozen base frequency: -0.006972 · 95% [-0.016805; 0.001499]. Below zero is better.

Separate hypothetical economics and comparisons

Selection	Known / unknown marks	Stress mark sum	Mean per selected entry	Mean per ALL opportunities
Prices only	245 / 0	$-10882.73	-20.14% · [-35.43%; -4.45%]	-9.46%
With volume	228 / 0	$-10739.71	-20.88% · [-36.21%; -5.41%]	-9.12%
No selection	522 / 0	$-24350.49	-22.14% · [-31.93%; -11.65%]	-22.14%

Change in all-opportunity mean versus «Prices only»: 0.34% · 95% [-1.10%; 1.64%].

Change in all-opportunity mean versus «No selection»: 13.02% · 95% [6.31%; 20.09%].

Skipped opportunities contribute zero, including zero fees; unknown selected marks never become zero. Positive selected marks alone do not prove superiority to either comparator.

Positive-mark criterion not passed.

Every ETF and quarter; descriptive only

Group	Dates	Δ Brier / prices	Δ Brier / base
SPY	174	-0.000001	-0.005997
QQQ	174	0.000148	-0.008202
IWM	174	0.000270	-0.006718
Q1	61	0.000145	-0.006379
Q2	62	0.000190	-0.016659
Q3	51	0.000069	0.004095
Q4	0	—	—

Calibration, sample requirements and concentration

Prices only

Probability bin	N	Forecast	Observed
0.00%–10.00%	25	7.44%	8.00%
10.00%–20.00%	289	15.55%	20.07%
20.00%–30.00%	171	23.95%	22.22%
30.00%–40.00%	36	32.65%	30.56%
40.00%–50.00%	1	40.43%	100.00%
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

With volume

Probability bin	N	Forecast	Observed
0.00%–10.00%	28	7.60%	10.71%
10.00%–20.00%	293	15.58%	19.45%
20.00%–30.00%	167	24.09%	23.35%
30.00%–40.00%	33	32.76%	30.30%
40.00%–50.00%	1	40.29%	100.00%
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Frozen base frequency

Probability bin	N	Forecast	Observed
0.00%–10.00%	0	—	—
10.00%–20.00%	522	14.64%	21.07%
20.00%–30.00%	0	—	—
30.00%–40.00%	0	—	—
40.00%–50.00%	0	—	—
50.00%–60.00%	0	—	—
60.00%–70.00%	0	—	—
70.00%–80.00%	0	—	—
80.00%–90.00%	0	—	—
90.00%–100.00%	0	—	—

Prices only · stress sum without the three best dates (diagnostic, never a filter): $-15175.73.

With volume · stress sum without the three best dates (diagnostic, never a filter): $-14758.71.

No selection · stress sum without the three best dates (diagnostic, never a filter): $-29246.49.

Source coverage and audit

Ledger / price-ready / volume-ready / forecasts: 522 / 522 / 522 / 522.

Extra exclusions due to volume history: 0; unknown future targets among forecasts: 0.

Registered 2026 dates: 174; independently checked morning windows including warmup: 594; minutes with fractional volume: 62875. Source files / scalar score checks: 2678 / 1044.

Independent audit SHA256 4c5f7ed7aca9902435ba67540317df058d53a4a74496ab22bcf5eb6a51293ad9

SHA256 383b113624256e687cddeccaaa4e300d96a3c4a41714025dc06377f198b3a8d2

H14-SX · far-option prices, expanded source · source calibration only; not an effectiveness result

SPY / QQQ / IWM. Missing preselected strikes were requested from Massive; source quality rules and the original H14 methodology were not relaxed. The earlier limited-universe experiment is preserved separately below.

2022–2024 calibration ready; annual effectiveness is not yet published.

ETF	Eligible dates	Required	Wing-score threshold	Credit threshold, % of stock
SPY	367	100	0.020000	0.2985%
QQQ	302	100	0.020000	0.3654%
IWM	128	100	0.030000	0.4590%

The wing score is not a probability or IV. These thresholds use prices only, no future-move labels; they do not prove that dangerous sessions can be predicted.

Coverage and limitations

Complete source rows: 798 / 2259. The final global session of 2024 is excluded from threshold calibration. Missing/stale/invalid quotes remain unknown, never zeros.

atm:quote_missing=1, complete=798, no_same_day_common_strike=338, short_session=18, source_missing=129, strike_offset_guard=4, wing:entry_quote_or_depth_unavailable=22, wing:entry_quote_or_depth_unavailable,quote_missing=1, wing:entry_quote_or_depth_unavailable,quote_stale=22, wing:quote_stale=926.

The eventual check uses a noon short ATM pair and a15:55 intrinsic liability mark, not buyback, assignment or annual return on funded capital. Previously explored years and nominal statistical intervals require later prospective confirmation.

SHA256 45c9143915c83f63a728de8c0ff67981e47dc1eceb4aae2e394e796aae7dfa19

SX · annual test of frozen price filters

SPY / QQQ / IWM, noon–15:55 ET. Alerts skip the short ATM call+put; observed distant contracts are not traded. Both avoided losses and forgone positive marks count.

2025

Insufficient evidence · 33 common dates / 99 ETF observations · 2 distinct severe-loss dates.

Filter	Trades / skips	Severe losses caught	Mean stress mark / W · 95% CI	Forgone positive stress marks
Wing-price filter	85 / 14	0 / 5 · 0.00%	-0.36% · [-31.95%; 25.16%]	$1709.00
Low-premium filter	80 / 19	0 / 5 · 0.00%	0.69% · [-28.31%; 24.15%]	$570.85
No filter	99 / 0	0 / 5 · 0.00%	1.16% · [-27.90%; 24.79%]	$0.00

Wing-filter change vs No filter: -1.52% · 95% [-6.14%; 3.75%].

Wing-filter change vs Low-premium filter: -1.05% · 95% [-5.86%; 4.33%].

Fixed gate, skips and normalization

Required: 100 common dates, 20 distinct severe-loss dates, 30 wing-alert dates, at least 950 defined precision and recall draws for BOTH filters, higher wing recall, and positive lower 95% bounds for BOTH paired utility differences. Current defined draws: 899/1000.

Filter	Alert dates	Precision · 95%	Recall · 95%	Avoided stress losses	Stress mark sum	Sum W: all / traded
Wing-price filter	11	0.00% · [0.00%; 0.00%]	0.00% · [—; —]	$456.49	$-3395.87	$31981.00 / $26658.00
Low-premium filter	7	0.00% · [0.00%; 0.00%]	0.00% · [—; —]	$496.20	$-2218.01	$31981.00 / $29519.00
No filter	0	— · [—; —]	0.00% · [—; —]	$0.00	$-2143.36	$31981.00 / $31981.00

W=B−4; B is the two-leg bid premium in dollars. Utility is the mean of each marked result/W over ALL eligible opportunities, with zero for skips—not a return on capital or the ratio of summed marks to summed W. Base cost=$4, stress total=$12 per traded pair. Severe loss means I>3B−12.

Every ETF and quarter · descriptive

ETF groups use their OWN available dates; quarters use the shared three-ETF cohort. These are not replacement primary tests.

Group	Dates / severe-loss dates	vs no filter · 95%	vs low premium · 95%
IWM	65 / 2	-0.16% · [-5.27%; 5.75%]	-1.20% · [-15.29%; 10.47%]
QQQ	161 / 6	0.74% · [-3.32%; 5.61%]	3.62% · [-1.36%; 8.56%]
SPY	128 / 6	-0.64% · [-4.50%; 3.14%]	1.04% · [-5.03%; 6.41%]
Q1	6 / 1	-3.53% · [-7.05%; 0.00%]	-2.23% · [-7.05%; 1.30%]
Q2	18 / 1	-5.08% · [-9.03%; -0.44%]	-5.08% · [-9.03%; -0.44%]
Q3	4 / 0	2.03% · [—; —]	7.45% · [—; —]
Q4	5 / 0	10.86% · [10.86%; 10.86%]	8.06% · [8.06%; 8.06%]

Coverage, ties and checksum

Ledger / saved scores / unknown outcomes among scores: 750 / 354 / 0.

IWM: 65 scores; distinct wing/credit 12/65; at threshold 7/0.

QQQ: 161 scores; distinct wing/credit 10/161; at threshold 61/0.

SPY: 128 scores; distinct wing/credit 9/128; at threshold 41/0.

Equal-to-threshold scores do not alert. A missing later target never erases its saved score.

Feature exclusions: atm:quote_stale: 1; complete: 354; short_session: 9; wing:entry_quote_or_depth_unavailable: 21; wing:entry_quote_or_depth_unavailable,quote_stale: 11; wing:quote_stale: 354;

Target availability: complete: 740; entry:atm:quote_stale: 1; entry:short_session: 9;

SHA256 c979d1126d4747a7e488a86a0fad7b183234127db304a17c1894b5f73215ad2f

2026

Insufficient evidence · 60 common dates / 180 ETF observations · 4 distinct severe-loss dates.

Filter	Trades / skips	Severe losses caught	Mean stress mark / W · 95% CI	Forgone positive stress marks
Wing-price filter	166 / 14	0 / 4 · 0.00%	-6.75% · [-27.85%; 13.67%]	$898.99
Low-premium filter	130 / 50	0 / 4 · 0.00%	-6.69% · [-27.14%; 12.82%]	$2099.27
No filter	180 / 0	0 / 4 · 0.00%	-5.51% · [-27.21%; 15.14%]	$0.00

Wing-filter change vs No filter: -1.24% · 95% [-4.37%; 1.41%].

Wing-filter change vs Low-premium filter: -0.06% · 95% [-5.49%; 6.10%].

Fixed gate, skips and normalization

Filter	Alert dates	Precision · 95%	Recall · 95%	Avoided stress losses	Stress mark sum	Sum W: all / traded
Wing-price filter	11	0.00% · [0.00%; 0.00%]	0.00% · [0.00%; 0.00%]	$244.00	$-3004.77	$45120.00 / $42299.00
Low-premium filter	23	0.00% · [0.00%; 0.00%]	0.00% · [0.00%; 0.00%]	$1768.51	$-2680.54	$45120.00 / $37590.00
No filter	0	— · [—; —]	0.00% · [0.00%; 0.00%]	$0.00	$-2349.78	$45120.00 / $45120.00

Every ETF and quarter · descriptive

ETF groups use their OWN available dates; quarters use the shared three-ETF cohort. These are not replacement primary tests.

Group	Dates / severe-loss dates	vs no filter · 95%	vs low premium · 95%
IWM	86 / 2	-2.11% · [-6.84%; 3.04%]	3.97% · [-4.42%; 13.22%]
QQQ	138 / 3	-0.02% · [-3.67%; 4.78%]	0.08% · [-4.58%; 4.81%]
SPY	132 / 1	-2.76% · [-5.74%; 0.25%]	0.91% · [-5.28%; 7.18%]
Q1	19 / 2	-1.46% · [-4.39%; 0.00%]	0.26% · [-2.32%; 2.84%]
Q2	28 / 2	-3.00% · [-7.85%; 0.14%]	-4.53% · [-12.10%; 3.15%]
Q3	13 / 0	2.88% · [-3.87%; 10.47%]	9.10% · [-4.98%; 21.42%]
Q4	0 / —	—	—

Coverage, ties and checksum

Ledger / saved scores / unknown outcomes among scores: 522 / 356 / 0.

IWM: 86 scores; distinct wing/credit 10/86; at threshold 6/0.

QQQ: 138 scores; distinct wing/credit 9/138; at threshold 43/0.

SPY: 132 scores; distinct wing/credit 6/132; at threshold 47/0.

Equal-to-threshold scores do not alert. A missing later target never erases its saved score.

Feature exclusions: complete: 356; wing:entry_quote_or_depth_unavailable: 14; wing:entry_quote_or_depth_unavailable,quote_stale: 7; wing:quote_stale: 145;

Target availability: complete: 522;

SHA256 624904dcfb42af9ffbb4c045a72e472a0e4b54809fe3df2ad803c377520b822f

Terminal intrinsic liability marks—not executed buybacks, assignment P&L or guaranteed profit. Omitted time value can overstate short results. The wing score is not a probability. Intervals are nominal 95%, 1000 circular five-date draws preserving ETFs together, without correction for repeated research. Previously explored years are not pristine holdouts; prospective confirmation is still required.

H14 · do distant option prices warn of large short losses?

SPY / QQQ / IWM, noon–15:55 ET. The proposed strategy sells one same-day ATM call and put; four wing contracts are observed, not traded. The protocol estimates thresholds from 2022–2024 prices and permits unchanged 2025/2026 tests only after all three ETFs meet the calibration requirements.

Evaluation rule: an alert skips the trade, counting both avoided losses and forgone wins. The wing-price score is not a probability or measured IV.

Insufficient calibration coverage. No 2025/2026 performance test was run. This is not a losing strategy result or zero predictive accuracy.

ETF	Valid calibration dates	Fixed minimum	Coverage
IWM	101	100	enough
QQQ	93	100	insufficient
SPY	139	100	enough

The joint rule requires all three ETFs. No symbol was dropped and no minimum was lowered.

Why coverage is missing

333 / 2259 ETF-date observations are feature-complete. The old archive collected only six neighboring strikes on each side at noon. A later-collected contract cannot replace a missing causal noon selection. Even a selected wing can have a genuinely stale quote; it is not made fresh by changing its label.

atm:quote_missing: 1

complete: 333

no_same_day_common_strike: 338

outside_causal_noon_collection_universe: 904

short_session: 18

source_missing: 129

strike_offset_guard: 4

wing:entry_quote_or_depth_unavailable: 17

wing:quote_stale: 515

Only prices were used for calibration, no future labels or profits. The first calculation was stopped for repeated annual-file decoding; its inputs and interruption record were retained. A cache-only correction preserved the calculations and produced this first completed result.

SHA256 1b8c79760bf00dc66c8ff5d162de375c0ec794048914e9a61733a01a51a59716

These are terminal intrinsic liability marks, not executed buybacks or assignment P&L. Omitted time value may overstate short results. Only fresh positive two-sided quotes within the causal noon collection universe are admitted; unavailable observations remain excluded. All years were previously accessed in other research and are not pristine holdouts. A historical pass would still need prospective verification.

H12 · avoiding rare large short-option losses · insufficient training evidence

SPY / QQQ / IWM, 12:00–15:55 ET, development2022–2024. Fresh same-day ATM quotes; current premium is already in the baseline. The candidate adds synchronized morning ETF moves and time-normalized morning-to-noon option-premium changes.

15 distinct tail dates among 484 eligible training dates; fixed minimum20. No model was fitted, no threshold was obtained, and2025/2026 were not evaluated. This is not zero accuracy or a negative strategy return.

Tail definition: I >3B−12, where B is the received bid premium in dollars, I is terminal intrinsic liability, W=B−4. A loss exceeds two hypothetical maximum wins W at that day’s credit. B>12 is required before entry.

Coverage and unchanged training gates

ETF-day ledger 2259; valid entries 1769; baseline-ready 1725; paired-feature-ready 1455. Final training after the purged session: 1452 / 484 rows / dates through 2024-12-30.

Source exclusions are not quiet-market labels. Missing source: 129; no same-day common ATM pair: 338; short-session rows: 18. The three ETFs on one date are not three independent tail events.

Validation quarter start	Training dates	Distinct tail dates / minimum	Result
2023-01-01	114	3 / 20	not trained
2023-04-01	153	4 / 20	not trained
2023-07-01	189	4 / 20	not trained
2023-10-01	225	5 / 20	not trained
2024-01-01	263	7 / 20	not trained
2024-04-01	302	7 / 20	not trained
2024-07-01	359	11 / 20	not trained
2024-10-01	422	13 / 20	not trained

All eight folds failed the predeclared minimum; no OOF probabilities or Brier score exist. We did not lower the minimum after seeing the data.

Terminal liability mark, not executable buyback, assignment or funded profit. Omitted time value can overstate short results. These periods were previously accessed by other studies; not pristine holdouts.

SHA256 a4cef2e5712f262a0d5f09b8cf44ae56d711cd0fb74272b7abe58e27d4063ca0

H10 · Can a nonlinear model identify movement beyond the premium?

SPY / QQQ / IWM, 12:00–15:55 New York. Models and alarm thresholds trained through 2024 and then frozen for both years. Inputs: past volatility, morning path, gap and current option cost. No news labels are used here.

Target is a terminal intrinsic-value mark—not executed profit. 100 × |price at 15:55 − ATM strike| minus the 12:00 call+put ask premium and $4 costs; stress costs $12. Remaining option time value, exercise/stock execution and their additional costs are not modeled.

2025 · advantage not confirmed

246 common dates / 738 ETF opportunities · 2025-01-02 — 2025-12-31. Δ Brier -0.00210 [-0.00662; 0.00249]. Negative is better than linear; the bracket is the 95% interval, in the same units.

Rule	Cases / dates	Precision / recall	Mean mark / premium	Stress mean · 95% CI	Sum of marks / paid premiums
Every eligible day	738 / 246	31.8% / 100.0%	-17.6%	-22.4% [-30.7%; -13.3%]	$-23353.59 / $165064.00
Linear	98 / 61	39.8% / 16.6%	-7.9%	-12.6% [-32.7%; 10.2%]	$-376.84 / $22232.00
Nonlinear	100 / 71	40.0% / 17.0%	-10.2%	-15.2% [-36.1%; 6.8%]	$-1529.63 / $19415.00

Probability reliability and tail diagnostics

Brier: constant 0.21852 · linear 0.21933 · nonlinear 0.21723.

N is the number of ETF observations on the same common dates. Mean predicted probability versus observed share of positive terminal marks. These bins describe reliability; no probability recalibration was fitted.

Probability bin	Constant frequency N · mean p / observed	Linear N · mean p / observed	Nonlinear N · mean p / observed
[0.0; 0.1)	0 · — / —	0 · — / —	0 · — / —
[0.1; 0.2)	0 · — / —	17 · 17.7% / 23.5%	15 · 17.0% / 20.0%
[0.2; 0.3)	246 · 28.8% / 36.2%	215 · 26.7% / 31.6%	235 · 27.0% / 30.6%
[0.3; 0.4)	492 · 33.1% / 29.7%	439 · 34.6% / 30.5%	432 · 34.3% / 31.0%
[0.4; 0.5)	0 · — / —	67 · 41.8% / 43.3%	54 · 42.2% / 46.3%
[0.5; 0.6)	0 · — / —	0 · — / —	2 · 51.3% / 50.0%
[0.6; 0.7)	0 · — / —	0 · — / —	0 · — / —
[0.7; 0.8)	0 · — / —	0 · — / —	0 · — / —
[0.8; 0.9)	0 · — / —	0 · — / —	0 · — / —
[0.9; 1.0]	0 · — / —	0 · — / —	0 · — / —

Rule	Worst base mark / premium	Large marks caught / missed	Best 3 dates: share of positive date marks	Stress sum without best 3 dates
Every eligible day	$-810.00 / -104.0%	47 / 0	—	—
Linear	$-567.00 / -100.9%	6 / 41	71.5%	$-6109.34 · 91 cases
Nonlinear	$-379.00 / -99.1%	6 / 41	64.1%	$-5417.63 · 92 cases

Large means base net mark / premium ≥100%, not annual capital return. The dollar and percentage minima can occur in different trades. Best-date removal is hindsight diagnosis only: it does not change either policy, headline result or confirmation gate. Concentration divides by the sum of positive date marks, not total net marks.

Linear · best dates (stress marks): 2025-04-08 $3303.00; 2025-06-05 $1014.50; 2025-05-21 $631.00.

Nonlinear · best dates (stress marks): 2025-05-21 $1243.00; 2025-06-05 $1014.50; 2025-07-31 $830.50.

ETF and quarter breakdowns, sample coverage

Group	Dates	Nonlinear alarms	Precision / recall	Stress mean · 95% CI
IWM	246	48	37.5% / 20.2%	-26.9% [-43.4%; -9.1%]
QQQ	247	26	38.5% / 14.5%	-9.4% [-46.9%; 31.5%]
SPY	247	28	42.9% / 15.6%	-2.9% [-36.1%; 37.6%]
Q1	59	16	50.0% / 13.8%	-2.3% [-39.8%; 32.9%]
Q2	62	25	32.0% / 15.1%	-3.8% [-64.7%; 65.7%]
Q3	63	28	39.3% / 17.5%	-21.2% [-57.7%; 21.5%]
Q4	62	31	41.9% / 21.3%	-25.6% [-48.6%; -4.1%]

Each ETF uses its own eligible dates; its sums need not add to the common-date headline. Quarters use the same common-date cohort.

Source rows 750 · forecasts 740 · unknown outcomes among forecasts 0. Missing prices and shortened sessions are not fabricated as zero outcomes.

SHA256 14c243f0bcea47a93909f6dcdcaaaf08e8b4e304f97c5002b88e882f3ce8b4b4

2026 · advantage not confirmed

174 common dates / 522 ETF opportunities · 2026-01-02 — 2026-09-11. Δ Brier -0.00090 [-0.00655; 0.00529]. Negative is better than linear; the bracket is the 95% interval, in the same units.

Rule	Cases / dates	Precision / recall	Mean mark / premium	Stress mean · 95% CI	Sum of marks / paid premiums
Every eligible day	522 / 174	32.8% / 100.0%	-17.8%	-22.1% [-31.9%; -11.7%]	$-20174.49 / $120812.00
Linear	82 / 50	36.6% / 17.5%	-8.7%	-12.8% [-34.5%; 9.1%]	$-517.49 / $20039.00
Nonlinear	91 / 65	35.2% / 18.7%	-18.1%	-23.2% [-40.4%; -5.8%]	$-2685.50 / $17776.00

Probability reliability and tail diagnostics

Brier: constant 0.22009 · linear 0.22311 · nonlinear 0.22220.

Probability bin	Constant frequency N · mean p / observed	Linear N · mean p / observed	Nonlinear N · mean p / observed
[0.0; 0.1)	0 · — / —	0 · — / —	0 · — / —
[0.1; 0.2)	0 · — / —	3 · 19.6% / 0.0%	10 · 16.9% / 20.0%
[0.2; 0.3)	174 · 28.8% / 31.0%	134 · 27.1% / 37.3%	144 · 26.7% / 34.7%
[0.3; 0.4)	348 · 33.1% / 33.6%	330 · 34.8% / 30.6%	317 · 34.4% / 31.5%
[0.4; 0.5)	0 · — / —	55 · 41.7% / 36.4%	51 · 42.2% / 37.3%
[0.5; 0.6)	0 · — / —	0 · — / —	0 · — / —
[0.6; 0.7)	0 · — / —	0 · — / —	0 · — / —
[0.7; 0.8)	0 · — / —	0 · — / —	0 · — / —
[0.8; 0.9)	0 · — / —	0 · — / —	0 · — / —
[0.9; 1.0]	0 · — / —	0 · — / —	0 · — / —

Rule	Worst base mark / premium	Large marks caught / missed	Best 3 dates: share of positive date marks	Stress sum without best 3 dates
Every eligible day	$-566.50 / -103.8%	41 / 0	—	—
Linear	$-465.00 / -100.7%	10 / 31	60.9%	$-4235.99 · 75 cases
Nonlinear	$-282.00 / -103.8%	9 / 32	46.3%	$-5029.00 · 85 cases

Linear · best dates (stress marks): 2026-06-11 $1639.00; 2026-03-31 $910.50; 2026-01-15 $513.00.

Nonlinear · best dates (stress marks): 2026-06-11 $774.00; 2026-01-15 $440.00; 2026-03-26 $401.50.

ETF and quarter breakdowns, sample coverage

Group	Dates	Nonlinear alarms	Precision / recall	Stress mean · 95% CI
IWM	174	47	31.9% / 27.8%	-32.1% [-53.4%; -10.0%]
QQQ	174	18	33.3% / 10.9%	-15.6% [-47.9%; 22.9%]
SPY	174	26	42.3% / 17.7%	-12.1% [-39.8%; 16.6%]
Q1	61	25	40.0% / 17.2%	-2.3% [-37.5%; 45.7%]
Q2	62	35	37.1% / 19.7%	-21.0% [-45.4%; 2.3%]
Q3	51	31	29.0% / 19.1%	-42.4% [-64.8%; -18.4%]
Q4	0	No evaluated dates—not zero return

Each ETF uses its own eligible dates; its sums need not add to the common-date headline. Quarters use the same common-date cohort.

Source rows 522 · forecasts 522 · unknown outcomes among forecasts 0. Missing prices and shortened sessions are not fabricated as zero outcomes.

SHA256 f1858bfb7569084a7dc5721bc65ff516b5153a7415fef9274b2bb5e34539e022

Confirmation requires a lower Brier than the linear model AND positive stress mean, both with the fixed 95% bounds, sufficient dates and signals. Date-block bootstrap: five dates, 1,000 draws. These years were previously studied; intervals are nominal within H10, not adjusted for the entire research search. Positive subgroup results cannot rescue a failed primary test.

Fixed minimums in each year separately: 100 common dates, 30 dates with a nonlinear alarm, 30 positive ETF outcomes and 950 defined conditional bootstrap draws. Alarm threshold: the 80th percentile of earlier out-of-fold scores for that ETF; it does not guarantee a 20% future selection rate.

Percentages are marks divided by option premiums, not annual returns on capital. No direction forecast, executable-profit guarantee or live-trading validation is claimed.

H11 · individual unresolved claims · text check failed

Fresh fixed 200-input sample from 2025: 196 usable, 4 source-unknown. 413 attempts for 393 requests, 38 failed attempts, including invalid model answers.

Independent blinded agent review: 21 positive / 68 negative / 11 unknown. Not a human gold standard or population sample.

Precision 85.7% [48.7%; 97.4%]; recall 28.6% [13.8%; 50.0%]. Wilson 95% intervals. Definite coverage 69/89 (77.5%).

Text quality only—not a volatility forecast. No market model was fitted with these labels.

Stages, confusion counts and limitations

35B extracts up to three literal actions; 27B conditionally checks each supplied action and its open state. These are not two independent extractions. With zero extracted claims, 27B checks absence only and cannot add a positive.

Extracted claims 50; zero-claim extractions 147; overflow / unclear selection 1 / 0. Zero extracted claims on reviewer-positive inputs: 5.

Sample	TP / FP / FN / TN	Abstain: positive / negative	Precision / recall
All 100	6 / 1 / 15 / 60	13 / 7	85.7% / 28.6%
Lexically selected 50	4 / 1 / 12 / 20	10 / 6	80.0% / 25.0%
Hash-selected controls 50	2 / 0 / 3 / 40	3 / 1	100.0% / 40.0%

FN already includes positive abstentions. Model positives on reviewer-unknown inputs: 1 — not confirmed successes.

All 200 model decisions: positive / negative / unknown 17 / 134 / 49. Fixed gate: ≥20 usable positive and negative review labels each, ≥90% coverage, ≥90% precision, ≥80% recall. No resampling after outputs.

The label describes an explicit unresolved systemic action in the retained excerpt at its declared version time. It does not establish novelty, active risk at the forecast cutoff, complete event inventory, historical availability, direction or profitability. Literal offsets identify retained input text, not semantic entailment or coordinates in the full archive. Semantic duplicates may remain.

Version, parsed-text and raw-layout references were frozen before any body reads. All 200 source inputs replayed against archived raw data. Blind labels froze before model calls.

SHA256 15f26216ce910083ebcb94902b129c90ef87335d96d73815400d07556eb20a57

Unresolved systemic action · text quality · extraction check failed

2025: fixed 200 inputs · 198 usable, 2 source-unknown. Calls 396, technical errors 0. This is text classification, not a volatility forecast.

Blinded independent agent review: 100 · 21 positive, 70 negative, 9 unknown. Source-usable definite labels: 91.

Positive precision 60.0% [23.1%; 88.2%]; recall 14.3% [5.0%; 34.6%]. Wilson 95% intervals; enriched agent-annotated sample, not population accuracy.

Definite model decisions on usable definite review labels: 76/91 (83.5%). Model-positive but reviewer-unknown: 2; these are not confirmed positives.

No market model has been fitted with these labels. The label concerns an explicit open claim in the retained excerpt at its declared version time—not a currently active risk, proven novelty, an event count, or future direction.

Counts, sampling strata and limitations

Sample	TP / FP / FN / TN	Abstain on positive / negative	Precision / recall	Overall admission
All 100	3 / 2 / 18 / 65	12 / 3	60.0% / 14.3%	80.0%
Lexically enriched 50	2 / 2 / 12 / 22	8 / 3	50.0% / 14.3%	68.0%
Hash-selected controls 50	1 / 0 / 6 / 43	4 / 0	100.0% / 14.3%	92.0%

FN includes abstentions on positive labels; the adjacent abstention column is a subset, not additional misses.

All 200 model labels: 14 positive, 154 negative, 32 unknown. The fixed gate requires 20 positive and 20 negative usable review labels, 90% definite coverage, 90% precision and 80% recall. Abstention on a definite positive is a recall miss.

The sample is not a human gold standard. Both models saw the same literal input independently; matching action-evidence blocks does not prove matching claim meaning. A citation block can contain more than one paragraph. Mixed-claim evidence flags in the blinded review labels: 0. A negative excerpt does not prove absence elsewhere in the article.

Version IDs and parsed-text hashes were saved before reading bodies. Raw-layout references were saved only after capture; their separate immutable-version/first-fetch audit is preserved. Historical availability of the Yahoo version remains unverified.

Quality report · SHA256 1bddfc8ba871bb275ddecb7937201c42d1ac8e6ab86d614238656035e91c1a77

Source audit · SHA256 cc0ea41976e88f8a2a945dd831bd64b11117f9352300b01ba7ebee7aaf76d482

Independent check · 2026

H6 / H2

The morning H2 signal did not confirm predictive improvement on independent 2026 data under the preregistered criterion.

The single primary check is 09:35 ET, SPY + QQQ + IWM on common dates. News through 09:25; market and option inputs known by 09:35. Models and thresholds were trained on 2025 and frozen before scoring 2026.

09:35 ET · 2026-01-02 — 2026-09-11 · 105 dates / 315 ETF observations. Positive events: 94 / 315 (29.8%).

Forecast	Brier ↓	Log loss ↓
Constant 2025 probability	0.2136	0.6207
Market + option cost	0.2136	0.6209
Market + cost + H2 news	0.2134	0.6192

ΔBrier (H2 − market) = -0.0003 · 95% CI [-0.0067; 0.0055].

Lower is better. CI: paired date-block bootstrap, five-session blocks, 1,000 draws. Confirmation requires the upper CI below zero and Brier better than the constant probability; noon and subgroup results cannot replace this criterion.

Frozen alerts	Precision	Recall	Alert coverage	Alerts / observations	TP / FP / FN	Alert dates
Market	34.2%	55.3%	48.3%	152 / 315	52 / 100 / 42	80 / 105
H2 + market	32.1%	63.8%	59.4%	187 / 315	60 / 127 / 34	78 / 105

Precision is the event share among alerts; recall is the share of events found; coverage is the share of observations flagged. Always answering “no” would have accuracy 70.2% — this is not model accuracy or evidence of usefulness.

Primary 09:35 calibration: ten probability bins

Each cell: N forecasts; mean predicted probability → observed event frequency. These are descriptive frequencies, not probability confidence intervals; small N is unstable. Dates may be dependent.

Probability bin	Constant probability	Market	H2 + market
[0.0%; 10.0%)	N=0— → —	N=0— → —	N=0— → —
[10.0%; 20.0%)	N=0— → —	N=2219.4% → 22.7%	N=4217.6% → 19.0%
[20.0%; 30.0%)	N=31523.3% → 29.8%	N=29323.5% → 30.4%	N=23924.7% → 33.5%
[30.0%; 40.0%)	N=0— → —	N=0— → —	N=3432.3% → 17.6%
[40.0%; 50.0%)	N=0— → —	N=0— → —	N=0— → —
[50.0%; 60.0%)	N=0— → —	N=0— → —	N=0— → —
[60.0%; 70.0%)	N=0— → —	N=0— → —	N=0— → —
[70.0%; 80.0%)	N=0— → —	N=0— → —	N=0— → —
[80.0%; 90.0%)	N=0— → —	N=0— → —	N=0— → —
[90.0%; 100.0%]	N=0— → —	N=0— → —	N=0— → —

Noon: a separate control, not the primary result

ET	Dates / observations	Events	Brier market → H2	ΔBrier · 95% CI	H2 precision / recallAlerts / all
12:00	105 / 315	24.8%	0.1850 → 0.1866	+0.0016[-0.0005; 0.0036]	30.6% / 28.2%72 / 315 · 22.9%

All six ETF × entry-time groups

ETF · ET	Dates / observations	Events	Brier market → H2	ΔBrier · 95% CI	H2 precision / recallAlerts / all
SPY · 09:35	105 / 105	27.6%	0.2002 → 0.2002	-0.0000[-0.0059; 0.0058]	30.6% / 51.7%49 / 105 · 46.7%
SPY · 12:00	105 / 105	24.8%	0.1853 → 0.1856	+0.0004[-0.0018; 0.0024]	23.3% / 26.9%30 / 105 · 28.6%
QQQ · 09:35	105 / 105	35.2%	0.2403 → 0.2391	-0.0012[-0.0085; 0.0063]	40.0% / 70.3%65 / 105 · 61.9%
QQQ · 12:00	105 / 105	26.7%	0.1983 → 0.2007	+0.0024[0.0000; 0.0054]	32.3% / 35.7%31 / 105 · 29.5%
IWM · 09:35	105 / 105	26.7%	0.2005 → 0.2009	+0.0004[-0.0079; 0.0069]	26.0% / 67.9%73 / 105 · 69.5%
IWM · 12:00	105 / 105	22.9%	0.1714 → 0.1733	+0.0019[-0.0006; 0.0042]	45.5% / 20.8%11 / 105 · 10.5%

Quarters: Q1 / Q2 / partial Q3

Quarter · ET	Dates / observations	Events	Brier market → H2	ΔBrier · 95% CI	H2 precision / recallAlerts / all
Q1 · 09:35	46 / 138	37.0%	0.2503 → 0.2447	-0.0056[-0.0162; 0.0002]	47.6% / 58.8%63 / 138 · 45.7%
Q2 · 09:35	26 / 78	25.6%	0.1886 → 0.1892	+0.0006[-0.0058; 0.0136]	30.8% / 60.0%39 / 78 · 50.0%
Q3 · 09:35	33 / 99	23.2%	0.1823 → 0.1888	+0.0064[-0.0048; 0.0195]	21.2% / 78.3%85 / 99 · 85.9%
Q1 · 12:00	46 / 138	31.9%	0.2249 → 0.2261	+0.0011[-0.0020; 0.0052]	38.5% / 34.1%39 / 138 · 28.3%
Q2 · 12:00	26 / 78	20.5%	0.1604 → 0.1609	+0.0005[-0.0019; 0.0044]	28.6% / 25.0%14 / 78 · 17.9%
Q3 · 12:00	33 / 99	18.2%	0.1487 → 0.1517	+0.0030[0.0005; 0.0068]	15.8% / 16.7%19 / 99 · 19.2%

Target: future minute realized volatility through 15:55 exceeds a Gaussian ATM proxy from the straddle price plus $4 cost. This is not observed IV, option profit or capital return. Yahoo text availability at the historical signal time is not verified: a historical holdout does not replace a future real-time test.

Forecasts: 630 / 1044 · unknown outcome: 0 · missing input features: 414. 2026-09-12 21:21:00 UTC

Option cost versus past volatility · 2024

The rule did not confirm a positive association under the predefined criterion.

One rule, no fitting: expected RV is the root mean square RV of the previous 10 trading sessions over the same clock interval. Alert if above the Gaussian proxy from today’s ATM call + put cost plus $4. Threshold 1 was not optimized. Primary entry09:35, horizon15:55 ET.

2024-01-02 — 2024-12-23 · 197 common dates · 591 ETF observations.

Entry ET	Dates / observations	Unfiltered event rate	Precision / recall	Alerts / coverage	Precision lift · 95% CI
09:35	197 / 591	17.8%	15.9% / 26.7%	176 · 29.8%	-1.9%[-8.3%; 5.2%]

TP / FP / FN: 28 / 148 / 77 · alert dates: 82. Lift is a difference in rates, in percentage points. Bootstrap: five-date blocks, 1,000 draws; undefined draws: 0.

Economic secondary: long straddle, exit15:55

Only common dates with fresh positive bids and size of at least one contract for both legs. Zero/stale bids are not replaced by sales. This secondary sample differs from the RV test. 47 dates / 141 observations; unknown exits among forecasts: 275 / 647.

Policy	Trades	Mean P&L, $	Mean net / debit	With $12 costs	Profitable
Unfiltered	141	-160.55	-59.3%[-64.5%; -54.3%]	-62.7%[-67.8%; -57.7%]	2.8%
On alert	30	-138.77	-63.4%[-72.4%; -53.7%]	-67.3%[-76.9%; -57.3%]	3.3%

One pair per instrument and entry. Percentages are mean return on paid premiums, not annual capital yield. Same-day ETFs are dependent; rows are not an investment portfolio.

Noon, all ETFs and quarters — no winner selection

Group	Dates / observations	Base rate	Precision / recall	Alerts / coverage	Lift · 95% CI
12:00	197 / 591	14.2%	8.4% / 17.9%	178 · 30.1%	-5.8%[-10.7%; 0.1%]
IWM · 09:35	197 / 197	18.3%	22.2% / 38.9%	63 · 32.0%	3.9%[-4.8%; 13.0%]
IWM · 12:00	197 / 197	14.2%	10.1% / 25.0%	69 · 35.0%	-4.1%[-10.5%; 3.9%]
QQQ · 09:35	225 / 225	17.8%	15.1% / 20.0%	53 · 23.6%	-2.7%[-10.3%; 6.0%]
QQQ · 12:00	225 / 225	14.2%	7.1% / 12.5%	56 · 24.9%	-7.1%[-12.7%; -0.1%]
SPY · 09:35	225 / 225	17.8%	8.6% / 15.0%	70 · 31.1%	-9.2%[-16.5%; -2.5%]
SPY · 12:00	225 / 225	12.9%	9.1% / 20.7%	66 · 29.3%	-3.8%[-9.5%; 3.0%]
Q1 · 09:35	39 / 117	17.9%	18.4% / 33.3%	38 · 32.5%	0.5%[-9.1%; 10.6%]
Q2 · 09:35	57 / 171	13.5%	4.9% / 8.7%	41 · 24.0%	-8.6%[-19.3%; 3.4%]
Q3 · 09:35	53 / 159	17.6%	9.3% / 17.9%	54 · 34.0%	-8.4%[-18.2%; 0.6%]
Q4 · 09:35	48 / 144	22.9%	32.6% / 42.4%	43 · 29.9%	9.6%[-8.7%; 25.4%]
Q1 · 12:00	39 / 117	9.4%	2.7% / 9.1%	37 · 31.6%	-6.7%[-13.7%; -0.4%]
Q2 · 12:00	57 / 171	19.9%	5.5% / 8.8%	55 · 32.2%	-14.4%[-25.9%; -3.0%]
Q3 · 12:00	53 / 159	10.1%	9.1% / 31.2%	55 · 34.6%	-1.0%[-7.9%; 7.5%]
Q4 · 12:00	48 / 144	16.0%	19.4% / 26.1%	31 · 21.5%	3.4%[-15.3%; 24.1%]

Fixed-rule2024 check, selected by source coverage, not outcomes. A short or incomplete session among the previous10 makes the forecast unknown; it is not replaced by an older day. RV exceeding a cost-based approximation is not observed IV and does not guarantee option profit.

Forecasts: 1294 / 1512 · unknown outcomes among forecasts: 0. 2026-09-12 21:47:11 UTC · SHA256 0e333e2fb2e251f826928e891c8663b3df6fa9fadeedb8f2d4214525a422fe02

Scheduled FOMC · cost-aware volatility check

2022–2023

Fixed primary: entry at 12:00 ET, FOMC decision days versus other Wednesdays, SPY/QQQ/IWM. This is a preannounced official-calendar rule, not an LLM/news-text signal.

Association not confirmed; fixed evidence requirements are not met. 15 FOMC dates / 79 ordinary Wednesdays; 282 ETF observations, not independent event dates.

Cost-exceeding ETF cases: 3/45 on FOMC dates, 22/237 on ordinary Wednesdays. Too few cost-exceeding event cases for the fixed 20-case requirement; this is not missing calendar data.

RV exceeds the cost proxy: 6.7% on FOMC / 9.3% on other Wednesdays. Δ -2.6 pp; 95% [-12.2; 7.5] pp; 99% [-14.7; 11.2] pp.

This is not actual IV, a probability forecast, option profit or 70–80% prediction accuracy. Fifteen potentially covered decisions are a small event sample.

All ETF, year and time comparisons

Group / control	Event / control dates	Event / control rate	Lift / intervals	Recall	5-min RV sensitivity
PRIMARY · 12:00Other Wednesdays	15 / 7945 / 237 ETF rows	6.7% / 9.3%	-2.6 pp95% [-12.2; 7.5] pp 99% [-14.7; 11.2] pp	12.0%22 missed ETF cases	24.4% / 11.4%Δ 13.1 pp
Morning control · 09:35Other Wednesdays	15 / 7745 / 231 ETF rows	6.7% / 11.3%	-4.6 pp95% [-13.0; 6.6] pp 99% [-14.5; 12.7] pp	10.3%26 missed ETF cases	13.3% / 11.3%Δ 2.1 pp
Secondary weekdays · 12:00All non-FOMC days	15 / 26545 / 795 ETF rows	6.7% / 9.1%	-2.4 pp95% [-9.7; 7.3] pp 99% [-10.7; 9.8] pp	4.0%72 missed ETF cases	24.4% / 9.9%Δ 14.5 pp
Morning + all days · 09:35All non-FOMC days	15 / 26245 / 786 ETF rows	6.7% / 10.7%	-4.0 pp95% [-11.7; 6.0] pp 99% [-13.2; 9.1] pp	3.4%84 missed ETF cases	13.3% / 11.5%Δ 1.9 pp
QQQ · 12:00Other Wednesdays	15 / 7915 / 79 ETF rows	6.7% / 10.1%	-3.5 pp95% [-15.4; 12.5] pp 99% [-18.1; 19.1] pp	11.1%8 missed ETF cases	13.3% / 12.7%Δ 0.7 pp
QQQ · 09:35Other Wednesdays	15 / 7715 / 77 ETF rows	6.7% / 11.7%	-5.0 pp95% [-16.5; 12.5] pp 99% [-18.6; 21.5] pp	10.0%9 missed ETF cases	6.7% / 13.0%Δ -6.3 pp
SPY · 12:00Other Wednesdays	15 / 7915 / 79 ETF rows	6.7% / 11.4%	-4.7 pp95% [-18.2; 12.4] pp 99% [-22.1; 19.0] pp	10.0%9 missed ETF cases	33.3% / 13.9%Δ 19.4 pp
SPY · 09:35Other Wednesdays	15 / 7815 / 78 ETF rows	6.7% / 15.4%	-8.7 pp95% [-22.5; 8.8] pp 99% [-26.0; 15.9] pp	7.7%12 missed ETF cases	20.0% / 11.5%Δ 8.5 pp
IWM · 12:00Other Wednesdays	15 / 7915 / 79 ETF rows	6.7% / 6.3%	0.3 pp95% [-8.9; 13.7] pp 99% [-10.4; 17.8] pp	16.7%5 missed ETF cases	26.7% / 7.6%Δ 19.1 pp
IWM · 09:35Other Wednesdays	15 / 7815 / 78 ETF rows	6.7% / 6.4%	0.3 pp95% [-7.6; 10.9] pp 99% [-9.1; 13.8] pp	16.7%5 missed ETF cases	13.3% / 9.0%Δ 4.4 pp
2022 · 12:00Other Wednesdays	7 / 3621 / 108 ETF rows	9.5% / 6.5%	3.0 pp95% [-11.1; 25.7] pp 99% [-12.8; 36.5] pp	22.2%7 missed ETF cases	42.9% / 13.9%Δ 29.0 pp
2023 · 12:00Other Wednesdays	8 / 4324 / 129 ETF rows	4.2% / 11.6%	-7.5 pp95% [-17.8; 1.9] pp 99% [-21.0; 5.2] pp	6.2%15 missed ETF cases	8.3% / 9.3%Δ -1.0 pp
2022 · 09:35Other Wednesdays	7 / 3521 / 105 ETF rows	9.5% / 6.7%	2.9 pp95% [-10.2; 24.8] pp 99% [-12.3; 31.4] pp	22.2%7 missed ETF cases	19.0% / 4.8%Δ 14.3 pp
2023 · 09:35Other Wednesdays	8 / 4224 / 126 ETF rows	4.2% / 15.1%	-10.9 pp95% [-18.1; -4.0] pp 99% [-20.7; -2.5] pp	5.0%19 missed ETF cases	8.3% / 16.7%Δ -8.3 pp

Rates concern the RV label only. The 09:35, all-weekday and 5-minute checks are secondary and cannot replace the noon Wednesday result. The 5-minute label retains exactly the same complete-minute sample.

Every scheduled date · including missing data

Date / entry ET	QQQ	SPY	IWM
2022-01-2612:00 → 15:55	RV/cost unknownStrict exit unknown	RV/cost unknownStrict exit unknown	RV/cost unknownStrict exit unknown
2022-01-2609:35 → 15:55	RV/cost unknownStrict exit unknown	RV/cost unknownStrict exit unknown	RV/cost unknownStrict exit unknown
2022-03-1612:00 → 15:55	Not above cost proxyRV 1.63% / cost proxy 1.69%RV5 1.82%Secondary net $-241.00 / -53.7%	Not above cost proxyRV 1.16% / cost proxy 1.24%RV5 1.33%Secondary net $-290.00 / -68.2%	Not above cost proxyRV 1.33% / cost proxy 1.57%RV5 1.56%Secondary net $-92.00 / -37.4%
2022-03-1609:35 → 15:55	Not above cost proxyRV 1.85% / cost proxy 2.13%RV5 2.02%Strict exit unknown	Not above cost proxyRV 1.35% / cost proxy 1.48%RV5 1.52%Secondary net $-196.00 / -39.0%	Not above cost proxyRV 1.58% / cost proxy 1.86%RV5 1.78%Strict exit unknown
2022-05-0412:00 → 15:55	Not above cost proxyRV 2.23% / cost proxy 2.37%RV5 2.28%Strict exit unknown	Not above cost proxyRV 1.57% / cost proxy 1.69%RV5 1.74%Strict exit unknown	Not above cost proxyRV 1.72% / cost proxy 1.89%RV5 2.16%Strict exit unknown
2022-05-0409:35 → 15:55	Not above cost proxyRV 2.40% / cost proxy 2.78%RV5 2.42%Strict exit unknown	Not above cost proxyRV 1.67% / cost proxy 2.07%RV5 1.83%Strict exit unknown	Not above cost proxyRV 1.89% / cost proxy 2.28%RV5 2.32%Strict exit unknown
2022-06-1512:00 → 15:55	Not above cost proxyRV 2.50% / cost proxy 2.79%RV5 2.73%Secondary net $-428.00 / -69.0%	Not above cost proxyRV 2.02% / cost proxy 2.17%RV5 2.27%Secondary net $-460.00 / -70.8%	Not above cost proxyRV 2.16% / cost proxy 2.57%RV5 2.57%Secondary net $-254.00 / -72.8%
2022-06-1509:35 → 15:55	Not above cost proxyRV 2.64% / cost proxy 3.25%RV5 2.84%Secondary net $-474.00 / -65.8%	Not above cost proxyRV 2.11% / cost proxy 2.51%RV5 2.34%Secondary net $-561.00 / -74.7%	Not above cost proxyRV 2.31% / cost proxy 3.07%RV5 2.63%Secondary net $-321.00 / -77.2%
2022-07-2712:00 → 15:55	Not above cost proxyRV 1.22% / cost proxy 1.94%RV5 1.39%Secondary net $60.00 / 12.9%	Not above cost proxyRV 1.00% / cost proxy 1.41%RV5 1.09%Strict exit unknown	Not above cost proxyRV 1.04% / cost proxy 1.58%RV5 1.28%Secondary net $32.00 / 14.3%
2022-07-2709:35 → 15:55	Not above cost proxyRV 1.36% / cost proxy 2.24%RV5 1.56%Strict exit unknown	Not above cost proxyRV 1.07% / cost proxy 1.58%RV5 1.16%Strict exit unknown	Not above cost proxyRV 1.22% / cost proxy 1.99%RV5 1.37%Strict exit unknown
2022-09-2112:00 → 15:55	Not above cost proxyRV 2.47% / cost proxy 2.68%RV5 2.99%Secondary net $-94.00 / -15.3%	Not above cost proxyRV 2.04% / cost proxy 2.11%RV5 2.49%Strict exit unknown	Not above cost proxyRV 1.96% / cost proxy 2.51%RV5 2.45%Strict exit unknown
2022-09-2109:35 → 15:55	Not above cost proxyRV 2.55% / cost proxy 3.00%RV5 3.05%Secondary net $-264.00 / -38.4%	Not above cost proxyRV 2.10% / cost proxy 2.31%RV5 2.54%Strict exit unknown	Not above cost proxyRV 2.07% / cost proxy 2.77%RV5 2.53%Secondary net $-202.00 / -51.8%
2022-11-0212:00 → 15:55	Above cost proxyRV 2.69% / cost proxy 2.49%RV5 2.42%Strict exit unknown	Above cost proxyRV 2.10% / cost proxy 1.96%RV5 1.92%Strict exit unknown	Not above cost proxyRV 2.20% / cost proxy 2.32%RV5 2.38%Strict exit unknown
2022-11-0209:35 → 15:55	Above cost proxyRV 2.75% / cost proxy 2.63%RV5 2.47%Strict exit unknown	Above cost proxyRV 2.13% / cost proxy 1.97%RV5 1.94%Strict exit unknown	Not above cost proxyRV 2.30% / cost proxy 2.51%RV5 2.45%Strict exit unknown
2022-12-1412:00 → 15:55	Not above cost proxyRV 2.27% / cost proxy 2.80%RV5 2.44%Secondary net $-268.00 / -41.6%	Not above cost proxyRV 1.81% / cost proxy 2.24%RV5 1.87%Secondary net $-361.00 / -50.3%	Not above cost proxyRV 2.12% / cost proxy 2.57%RV5 2.15%Secondary net $-246.00 / -66.5%
2022-12-1409:35 → 15:55	Not above cost proxyRV 2.34% / cost proxy 2.70%RV5 2.49%Secondary net $-437.00 / -70.6%	Not above cost proxyRV 1.86% / cost proxy 2.10%RV5 1.91%Secondary net $-413.00 / -61.5%	Not above cost proxyRV 2.20% / cost proxy 2.47%RV5 2.22%Secondary net $-230.00 / -65.0%
2023-02-0112:00 → 15:55	Not above cost proxyRV 1.62% / cost proxy 2.17%RV5 1.78%Strict exit unknown	Not above cost proxyRV 1.22% / cost proxy 1.59%RV5 1.36%Secondary net $8.00 / 1.6%	Not above cost proxyRV 1.42% / cost proxy 1.91%RV5 1.59%Strict exit unknown
2023-02-0109:35 → 15:55	Not above cost proxyRV 1.71% / cost proxy 2.30%RV5 1.85%Strict exit unknown	Not above cost proxyRV 1.29% / cost proxy 1.66%RV5 1.43%Secondary net $-13.00 / -2.4%	Not above cost proxyRV 1.57% / cost proxy 1.98%RV5 1.77%Strict exit unknown
2023-03-2212:00 → 15:55	Not above cost proxyRV 1.62% / cost proxy 2.05%RV5 1.56%Strict exit unknown	Not above cost proxyRV 1.32% / cost proxy 1.72%RV5 1.29%Strict exit unknown	Not above cost proxyRV 1.68% / cost proxy 2.18%RV5 1.66%Strict exit unknown
2023-03-2209:35 → 15:55	Not above cost proxyRV 1.68% / cost proxy 2.13%RV5 1.62%Secondary net $-270.00 / -51.7%	Not above cost proxyRV 1.36% / cost proxy 1.73%RV5 1.33%Secondary net $-120.00 / -21.9%	Not above cost proxyRV 1.80% / cost proxy 2.42%RV5 1.79%Strict exit unknown
2023-05-0312:00 → 15:55	Not above cost proxyRV 1.26% / cost proxy 1.43%RV5 0.78%Secondary net $-128.00 / -35.4%	Not above cost proxyRV 1.06% / cost proxy 1.22%RV5 0.63%Strict exit unknown	Not above cost proxyRV 1.42% / cost proxy 1.56%RV5 0.94%Secondary net $-143.00 / -67.5%
2023-05-0309:35 → 15:55	Not above cost proxyRV 1.35% / cost proxy 1.46%RV5 0.90%Secondary net $-135.00 / -36.6%	Not above cost proxyRV 1.13% / cost proxy 1.22%RV5 0.70%Strict exit unknown	Not above cost proxyRV 1.58% / cost proxy 1.70%RV5 1.15%Secondary net $-161.00 / -70.0%
2023-06-1412:00 → 15:55	Not above cost proxyRV 1.11% / cost proxy 1.38%RV5 1.13%Secondary net $-326.00 / -81.9%	Not above cost proxyRV 0.81% / cost proxy 0.91%RV5 0.83%Strict exit unknown	Not above cost proxyRV 1.11% / cost proxy 1.41%RV5 1.04%Strict exit unknown
2023-06-1409:35 → 15:55	Not above cost proxyRV 1.19% / cost proxy 1.44%RV5 1.21%Secondary net $-274.00 / -66.3%	Not above cost proxyRV 0.85% / cost proxy 0.96%RV5 0.86%Secondary net $-235.00 / -71.0%	Not above cost proxyRV 1.32% / cost proxy 1.59%RV5 1.16%Strict exit unknown
2023-07-2612:00 → 15:55	Not above cost proxyRV 0.80% / cost proxy 1.16%RV5 0.94%Secondary net $-220.00 / -63.6%	Not above cost proxyRV 0.55% / cost proxy 0.72%RV5 0.64%Secondary net $-193.00 / -75.1%	Not above cost proxyRV 0.64% / cost proxy 0.93%RV5 0.80%Secondary net $-106.00 / -74.6%
2023-07-2609:35 → 15:55	Not above cost proxyRV 0.93% / cost proxy 1.36%RV5 1.02%Secondary net $-278.00 / -68.8%	Not above cost proxyRV 0.59% / cost proxy 0.82%RV5 0.66%Secondary net $-230.00 / -78.2%	Not above cost proxyRV 0.85% / cost proxy 1.20%RV5 0.96%Secondary net $-148.00 / -80.4%
2023-09-2012:00 → 15:55	Not above cost proxyRV 0.88% / cost proxy 0.98%RV5 0.73%Strict exit unknown	Not above cost proxyRV 0.66% / cost proxy 0.73%RV5 0.59%Strict exit unknown	Not above cost proxyRV 0.71% / cost proxy 1.02%RV5 0.64%Strict exit unknown
2023-09-2009:35 → 15:55	Not above cost proxyRV 0.98% / cost proxy 1.14%RV5 0.84%Strict exit unknown	Not above cost proxyRV 0.71% / cost proxy 0.79%RV5 0.64%Strict exit unknown	Not above cost proxyRV 0.87% / cost proxy 1.15%RV5 0.74%Strict exit unknown
2023-11-0112:00 → 15:55	Not above cost proxyRV 0.93% / cost proxy 1.08%RV5 0.82%Secondary net $16.00 / 5.3%	Not above cost proxyRV 0.76% / cost proxy 0.90%RV5 0.66%Secondary net $-125.00 / -41.8%	Not above cost proxyRV 0.96% / cost proxy 1.09%RV5 0.82%Secondary net $-99.00 / -71.2%
2023-11-0109:35 → 15:55	Not above cost proxyRV 1.09% / cost proxy 1.36%RV5 1.01%Strict exit unknown	Not above cost proxyRV 0.87% / cost proxy 1.07%RV5 0.78%Strict exit unknown	Not above cost proxyRV 1.19% / cost proxy 1.41%RV5 1.02%Strict exit unknown
2023-12-1312:00 → 15:55	Not above cost proxyRV 0.72% / cost proxy 0.87%RV5 0.77%Strict exit unknown	Not above cost proxyRV 0.54% / cost proxy 0.69%RV5 0.69%Strict exit unknown	Above cost proxyRV 1.42% / cost proxy 1.25%RV5 1.81%Strict exit unknown
2023-12-1309:35 → 15:55	Not above cost proxyRV 0.78% / cost proxy 0.94%RV5 0.80%Strict exit unknown	Not above cost proxyRV 0.57% / cost proxy 0.69%RV5 0.70%Strict exit unknown	Above cost proxyRV 1.52% / cost proxy 1.36%RV5 1.87%Strict exit unknown

All 16 scheduled dates and both entries are retained. Jan 26, 2022 precedes available source indexes. Missing contracts, entries, minutes or exits are not losses or zero-volatility days.

Secondary economics · separate strict-exit sample

12:00 · Common strict-exit dates 5 / 43 (FOMC / other Wednesday). Mean net / paid premium -56.8% / -53.6%; stress -59.3% / -58.6%. Mean dollars per one pair $-224.47 / $-97.71. Unknown exit ETF rows 23 / 74.

09:35 · Common strict-exit dates 3 / 13 (FOMC / other Wednesday). Mean net / paid premium -71.4% / -61.5%; stress -73.4% / -64.7%. Mean dollars per one pair $-343.56 / $-181.44. Unknown exit ETF rows 26 / 125.

Buy one ATM call+put at asks; sell both at fresh positive 15:55 bids with size. Base costs $4, stress $12. This conditional sample may be selective and does not estimate all scheduled events. Percentages divide net by paid premium: not account yield or annual returns. No missing exit is replaced by zero or a floor.

Fixed criterion, sources and limitations

For the primary check: at least 80 common Wednesdays, 12 FOMC dates, 5 event dates in each year, 20 positive ETF cases, 950 defined bootstrap draws, positive lift in both years and overall 95% lower bound above zero. 99% is additional disclosure, not a replacement criterion. Primary circular blocks contain 5 observed Wednesdays; ETF rows from one date stay together. The all-days secondary control uses 5 observed trading days. Defined primary draws: 1000/1000.

Complete 1-minute closes from entry to 15:55 are compared with the Gaussian ATM cost proxy. Strikes, jumps, American exercise and remaining time limit that approximation. No fitted threshold, event-date optimization or direction rule. 2024–2026 were already examined elsewhere; December 2023 stock RV was previously used as warmup.

Calendar announcements: 2022 · 2023. 3006 source rows, 2096 complete entries, 2 unknown outcomes after entry.

SHA256 a0a6bbd80dbb2f7b73ffdd0db27ba1703d9cb5f87cebf51f6c26122520fc67d8

News meaning / timing check · insufficient evidence

2025 · Selected 200; usable inputs 197; need source review 3. Model calls 394, technical errors 0.

Blinded independent agent review: 50 · 47 negative, 3 unknown, 0 confirmed positive.

On those negatives: model agrees on 43; abstains on 4. Positive precision and recall cannot be measured here.

Later review of all 5 model-positive cases: 1 negative, 4 unknown, 0 supported inside the target interval.

Semantic gate not passed; no market fitting. This is not evidence that all news precursors are absent.

Limits, source defect and checksums

The test asks whether a specific systemic action is still pending inside the trading interval, not whether a story is important or bullish. The 200 inputs were fixed before model calls; the 50 labels were frozen before reviewers saw local-model outputs. This is an enriched sample and agent annotation, not human ground truth or population accuracy.

All model decisions: 173 negative, 5 positive, 22 unknown. The later review selected model-positive cases after outputs existed; it is not a new blind precision/recall estimate. A successful API response or literal citation does not prove the time inference.

Jan 29: the impossible “30 p.m.” and conflicting timing already occur in the captured Yahoo article body; this is not a parser-created error.

2 short-session contexts mistakenly extended to 15:55 instead of 13:00. A pre-call admission flag makes them unknown; original inputs and raw model decisions remain preserved. Historical availability of Yahoo text is unverified. Unknown does not mean no risk.

audit · SHA256 9555513892b9ffc39ba0be121607eb37e259a9c7fb2524fd0005426da605e22e

models · SHA256 1b871f2fcf601745b82b6622b071f4c8ce7dc4f824a4f4fab99331b89333c410

followup · SHA256 948edb729d293f7ff6f1d92040598c8da6a8844b26ec6f92835d460490922269

Model-filtered news · all registered checks

10/10

2025 development, not independent confirmation. H2m adds genre/topic-filtered news flow to each stated baseline. All ten attempts are retained; different targets and samples must not be ranked together.

All eight option checks have a negative mean after stress costs. Neither variance check confirms an improvement in forecast error.

All ten comparisons and their samples

Target / training	Baseline → +H2m	Out-of-fold sample	Native error: baseline → enhanced	Error delta / 95% interval	Policy result
Win · fixed 15:55	Market + premium→ + H2m	165 dates / 9902025-05-01 — 2025-12-31	Brier0.21732 → 0.22070	+0.003388[+0.000209; +0.007533]	Selected stress mean-30.37% → -30.42%+H2m [-45.12%; -16.25%]Selected 97 / 744; 70 datesPolicy period 2025-07-01 — 2025-12-31
Win · fixed 15:55	Market + premium + H2→ + H2m	165 dates / 9902025-05-01 — 2025-12-31	Brier0.21820 → 0.22105	+0.002848[-0.000153; +0.006873]	Selected stress mean-21.03% → -18.77%+H2m [-37.93%; -2.80%]Selected 120 / 744; 63 datesPolicy period 2025-07-01 — 2025-12-31
Win · TP25	Market + premium→ + H2m	165 dates / 9902025-05-01 — 2025-12-31	Brier0.25420 → 0.25556	+0.001359[-0.002776; +0.006207]	Selected stress mean-19.62% → -20.54%+H2m [-30.60%; -11.22%]Selected 151 / 744; 68 datesPolicy period 2025-07-01 — 2025-12-31
Win · TP25	Market + premium + H2→ + H2m	165 dates / 9902025-05-01 — 2025-12-31	Brier0.25466 → 0.25640	+0.001748[-0.002712; +0.007142]	Selected stress mean-12.56% → -15.73%+H2m [-26.06%; -8.20%]Selected 146 / 744; 62 datesPolicy period 2025-07-01 — 2025-12-31
Return magnitudeWeekly expanding	Market + premium→ + H2m	163 dates / 9782025-05-05 — 2025-12-31	MSE0.57269 → 0.59485	+0.022164[+0.005702; +0.040845]	Selected stress mean-33.02% → -23.27%+H2m [-36.23%; -10.94%]Selected 166 / 978; 76 datesPolicy period 2025-05-05 — 2025-12-31
Return magnitudeWeekly, last 90 sessions	Market + premium→ + H2m	163 dates / 9782025-05-05 — 2025-12-31	MSE0.59814 → 0.64262	+0.044480[+0.009214; +0.078050]	Selected stress mean-29.82% → -22.93%+H2m [-35.38%; -8.92%]Selected 202 / 978; 83 datesPolicy period 2025-05-05 — 2025-12-31
Return magnitudeWeekly expanding	Market + premium + H2→ + H2m	163 dates / 9782025-05-05 — 2025-12-31	MSE0.59329 → 0.61739	+0.024100[+0.005128; +0.049546]	Selected stress mean-24.13% → -27.41%+H2m [-40.16%; -11.14%]Selected 194 / 978; 84 datesPolicy period 2025-05-05 — 2025-12-31
Return magnitudeWeekly, last 90 sessions	Market + premium + H2→ + H2m	163 dates / 9782025-05-05 — 2025-12-31	MSE0.63429 → 0.67639	+0.042100[+0.011981; +0.070449]	Selected stress mean-31.05% → -28.39%+H2m [-38.71%; -12.69%]Selected 227 / 978; 92 datesPolicy period 2025-05-05 — 2025-12-31
Realised volatility	Market + premium→ + H2m	166 dates / 9962025-05-01 — 2025-12-31	Brier0.16748 → 0.16760	+0.000116[-0.001952; +0.001997]	Precision +22.39% → +27.91%Recall +8.88% → +21.30%Alarm coverage +8.93% → +17.20%750 policy observations; not option profitPolicy period 2025-07-01 — 2025-12-31
Realised volatility	Market + premium + H2→ + H2m	166 dates / 9962025-05-01 — 2025-12-31	Brier0.16616 → 0.16583	-0.000330[-0.002507; +0.001587]	Precision +31.39% → +27.38%Recall +25.44% → +27.22%Alarm coverage +18.27% → +22.40%750 policy observations; not option profitPolicy period 2025-07-01 — 2025-12-31

Negative error delta is better, but it is not the profit change. Brier tests binary outcomes; MSE tests the magnitude of net/debit. Policy samples exclude threshold warmup and can be smaller than score samples.

Units, limits and report identities

Fixed and TP25 purchases use the frozen conservative residual-write-off policy. Percentages are net P&L divided by paid option premiums, not account yield or guaranteed market liquidation. The two entry times are alternatives. A negative policy floor does not prove a negative actual liquidation.

Variance checks only future minute-return RV against a Gaussian ATM cost proxy. The proxy is not observed IV; realised volatility is not straddle profit. Precision and recall refer only to this label.

Intervals resample paired blocks of dates, keeping ETF/time observations together. A historical 80th-percentile threshold does not fix future alarms at20%. Q1/April were training; these semantic variants remain untested on2026. Historical Yahoo timing and model memory of2025 remain unverified. Genre/topic agreement with a reviewer is not forecast accuracy or proof of fresh ETF-relevant events.

Win · fixed 15:55 · SHA256 0a924313bfe17608b8e9ca92a5c5fa5faa5a5041bebeb3545728d483e5bdea4b

Win · TP25 · SHA256 0aedfc1369f5c693ea7a984698f47cf736337cbaad1e684a5c29676414bdeb65

Return magnitude · SHA256 e78d66d8f278469299fddf2d08c8f41e68e4d0742ff04ceade11ac16523c51d3

Realised volatility · SHA256 5657a07603b35acd64a6bb3b7a34ce67508634032fce507eeabb65c5de5b3488

Movement beyond the option premium

SPY · QQQ · IWM

2025 development · no validated signal. These option-profit variants have not been tested on 2026; the separate H6/H2 check is shown above. Historical news availability (PIT) is unverified.

1477 / 1500 eligible observationsEntry 09:35 / 12:00 → 15:55 NY

Conservative valuation: a stale penny residual is written off, not sold at an observed bid.

Data, exclusions and strict-cohort checks · 23 excluded / unknown

Can signals known before entry identify days when an ATM call + put earns more than both purchase premiums and costs? Entry at 09:35 or 12:00 New York, exit at 15:55; the two entry times are alternative strategies.

This is development, not a discovered signal. All of 2025 has already been used for research. These option-profit variants have no independent 2026 evaluation. Historical Yahoo availability is unverified; a positive development result is not proof of a live forecast.

Conservative residual-write-off valuation

A stale penny residual is valued at zero under a separately frozen abandonment / do-not-exercise policy. This is not an observed market bid-exit. A negative conservative estimate does not prove that liquidation at the actual market bid would lose money. The strict experiment is preserved separately; changed coverage is not proof of better predictions.

Source ETF / entry observations: 1500Joined observations: 1477Excluded / unknown: 23

Observations with a confirmed zero-bid write-off: 239; with a stale-residual write-off: 251. Categories may overlap; these are not necessarily executed sales.

Unknown option exits: 3. These observations are excluded, not assigned a zero. Missing exits can bias the comparison; the results describe only the admitted sample.

Coverage and exclusions

Reason	N
Unknown / stale exit	3
Unavailable entry	2
Early-close session	18
Other option-source exclusion	0
Unavailable market controls	0
Unavailable news coverage	0
Other exclusions	0

ETF / NY	Source rows	Option targets available	Unknown	Excluded
SPY · 09:35	250	246	1	3
SPY · 12:00	250	246	1	3
QQQ · 09:35	250	246	1	3
QQQ · 12:00	250	246	1	3
IWM · 09:35	250	247	0	3
IWM · 12:00	250	246	1	3

Strict / residual-policy comparison: 1226 common source observations; 251 additional admitted observations. On the common cohort, inputs, targets and separately refitted models agree exactly. Any difference in the expanded-cohort tables is not evidence that the valuation rule improved prediction.

Strict sensitivity: refit on the common source cohort

The table below uses only common source observations for both training and evaluation. The expanded-cohort models below are trained again on a larger sample; their metrics are not paired with this table.

Family	Rows / prediction dates / policy dates	Brier base → family	Δ Brier	Selected stress mean · family	Paired stress difference / opportunity
H1 · observed market path	1132 / 30 / 11	0.1789 → 0.1793	+0.0004 [-0.0010; +0.0019]	-23.45% [-47.82%; -4.14%]	+1.37 pp [-0.62 pp; +3.06 pp]
H2 · news tempo	1186 / 52 / 18	0.1736 → 0.1736	-0.0000 [-0.0017; +0.0015]	-38.50% [-54.13%; -21.06%]	-2.89 pp [-9.36 pp; +0.06 pp]
Reference · static headline topics	1226 / 78 / 52	0.1658 → 0.1718	+0.0061 [-0.0014; +0.0133]	-36.33% [-45.47%; -24.95%]	-0.73 pp [-2.53 pp; +0.76 pp]

Strict source SHA-256: 56a90635af9f6ec86f4669ee708da4f3f96ef2ce775390ad6469abd00768e2ed
Common refit SHA-256: 347c470bb94aa0701f1ac5b4d1e5b629fdeddd106c579dc9a1fdbcdbc7561d54

Each family is compared with market + option-cost controls on exactly its own eligible dates. Family coverage can differ. Lower Brier is better; Δ Brier = added features − matching baseline. This is prediction error, not percentage accuracy.

Baseline: market + premium, matched dates. Lower Brier is better; return is net / paid premium, not account yield.

H1 · observed market path

64 prediction dates · 44 policy dates

Metric	Baseline	+ features	Δ
Brier	0.2236	0.2243	+0.0007 [-0.0005; +0.0021]
Selections / opportunities	62 / 264 23.48%	56 / 264 21.21%	—
Selected mean · stress	-16.14% [-38.36%; +12.19%]	-13.77% [-38.60%; +17.02%]	—
Per opportunity · stress	Skipped = 0	Same cases	+0.87 pp [-1.78 pp; +3.50 pp]

Distinct selected dates, baseline / family: 36 / 36.

All six ETF / entry comparisons

ETF / NY	N	Brier base → family	Δ Brier	Coverage base → family	Stress mean base → family
SPY · 09:35	162	0.2382 → 0.2377	-0.0006	4.96% → 5.79% Selected 6 → 7	+9.71% → +7.72%
SPY · 12:00	165	0.2060 → 0.2065	+0.0005	0.00% → 0.00% Selected 0 → 0	— → —
QQQ · 09:35	157	0.2240 → 0.2266	+0.0025	8.55% → 2.56% Selected 10 → 3	-29.71% → -55.18%
QQQ · 12:00	165	0.1831 → 0.1830	-0.0001	7.26% → 8.87% Selected 9 → 11	-25.01% → -19.67%
IWM · 09:35	71	0.2441 → 0.2436	-0.0005	28.26% → 28.26% Selected 13 → 13	+15.97% → +11.42%
IWM · 12:00	166	0.2193 → 0.2191	-0.0001	34.40% → 32.00% Selected 43 → 40	-27.91% → -26.61%

Features and sample coverage

Efficiency and RV acceleration over the previous 30 clock minutes. At 09:35: 25 premarket + 5 regular-session minutes; at noon: entirely regular trading.

Feature-eligible rows: 1342; missing features: 135; common prediction dates: 64; policy dates: 44; unavailable group / fold fits: 1.

Sensitivity: exclude one calendar month

Excluded month	Selected stress mean · baseline	Selected stress mean · family
2025-09	-16.07%	-13.18%
2025-10	-14.54%	-16.13%
2025-11	-32.55%	-31.11%
2025-12	-3.44%	+3.82%

Calibration: forecasts and observed frequencies

Counts are ETF / entry observations, not independent trading dates. Sparse bins do not establish calibration.

Matching baseline

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	0	—	—
0.1–0.2	5	18.70%	20.00%
0.2–0.3	167	25.93%	33.53%
0.3–0.4	210	33.98%	33.33%
0.4–0.5	2	40.58%	50.00%
0.5–0.6	0	—	—
0.6–0.7	0	—	—
0.7–0.8	0	—	—
0.8–0.9	0	—	—
0.9–1.0	0	—	—

With this family

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	0	—	—
0.1–0.2	11	19.01%	27.27%
0.2–0.3	163	26.15%	33.13%
0.3–0.4	205	34.09%	34.15%
0.4–0.5	5	40.57%	20.00%
0.5–0.6	0	—	—
0.6–0.7	0	—	—
0.7–0.8	0	—	—
0.8–0.9	0	—	—
0.9–1.0	0	—	—

H2 · news tempo

165 prediction dates · 124 policy dates

Metric	Baseline	+ features	Δ
Brier	0.2173	0.2182	+0.0009 [-0.0020; +0.0040]
Selections / opportunities	71 / 744 9.54%	113 / 744 15.19%	—
Selected mean · stress	-30.37% [-43.33%; -16.35%]	-21.03% [-36.79%; -4.76%]	—
Per opportunity · stress	Skipped = 0	Same cases	-0.30 pp [-1.73 pp; +1.32 pp]

Distinct selected dates, baseline / family: 58 / 69.

All six ETF / entry comparisons

ETF / NY	N	Brier base → family	Δ Brier	Coverage base → family	Stress mean base → family
SPY · 09:35	166	0.2424 → 0.2452	+0.0029	0.00% → 7.20% Selected 0 → 9	— → -32.20%
SPY · 12:00	165	0.2046 → 0.2056	+0.0010	0.00% → 2.42% Selected 0 → 3	— → -48.10%
QQQ · 09:35	165	0.2223 → 0.2225	+0.0001	4.03% → 15.32% Selected 5 → 19	-34.23% → -0.44%
QQQ · 12:00	165	0.1828 → 0.1843	+0.0015	10.48% → 15.32% Selected 13 → 19	-26.37% → -22.74%
IWM · 09:35	166	0.2331 → 0.2324	-0.0008	0.80% → 6.40% Selected 1 → 8	+56.57% → +25.23%
IWM · 12:00	166	0.2203 → 0.2214	+0.0011	42.40% → 44.80% Selected 53 → 56	-32.49% → -30.84%

Features and sample coverage

Risk-topic concentration, recent 2-hour news burst versus the preceding 22 hours, and changes relative to earlier same-time windows. Lexical markers are not semantic novelty or LLM assessments.

Feature-eligible rows: 1417; missing features: 60; common prediction dates: 165; policy dates: 124; unavailable group / fold fits: 0.

Sensitivity: exclude one calendar month

Excluded month	Selected stress mean · baseline	Selected stress mean · family
2025-07	-30.37%	-21.03%
2025-08	-30.37%	-21.95%
2025-09	-28.33%	-19.37%
2025-10	-32.48%	-18.41%
2025-11	-37.94%	-35.43%
2025-12	-23.67%	-13.63%

Calibration: forecasts and observed frequencies

Counts are ETF / entry observations, not independent trading dates. Sparse bins do not establish calibration.

Matching baseline

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	13	7.56%	30.77%
0.1–0.2	64	17.17%	25.00%
0.2–0.3	347	25.68%	27.95%
0.3–0.4	447	34.12%	34.00%
0.4–0.5	95	43.21%	30.53%
0.5–0.6	24	53.90%	29.17%
0.6–0.7	0	—	—
0.7–0.8	0	—	—
0.8–0.9	0	—	—
0.9–1.0	0	—	—

With this family

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	12	7.14%	25.00%
0.1–0.2	76	17.15%	27.63%
0.2–0.3	390	25.77%	29.74%
0.3–0.4	386	34.37%	32.38%
0.4–0.5	111	42.54%	33.33%
0.5–0.6	12	54.25%	25.00%
0.6–0.7	3	61.23%	0.00%
0.7–0.8	0	—	—
0.8–0.9	0	—	—
0.9–1.0	0	—	—

Reference · static headline topics

165 prediction dates · 124 policy dates

Metric	Baseline	+ features	Δ
Brier	0.2167	0.2234	+0.0067 [+0.0002; +0.0128]
Selections / opportunities	72 / 744 9.68%	70 / 744 9.41%	—
Selected mean · stress	-27.75% [-41.25%; -10.92%]	-22.34% [-43.70%; -1.89%]	—
Per opportunity · stress	Skipped = 0	Same cases	+0.58 pp [-0.41 pp; +1.65 pp]

Distinct selected dates, baseline / family: 51 / 50.

All six ETF / entry comparisons

ETF / NY	N	Brier base → family	Δ Brier	Coverage base → family	Stress mean base → family
SPY · 09:35	166	0.2401 → 0.2499	+0.0098	3.20% → 4.00% Selected 4 → 5	-59.18% → -31.05%
SPY · 12:00	165	0.2060 → 0.2121	+0.0061	0.00% → 0.81% Selected 0 → 1	— → -85.45%
QQQ · 09:35	165	0.2215 → 0.2349	+0.0135	12.10% → 5.65% Selected 15 → 7	-30.45% → -9.03%
QQQ · 12:00	165	0.1831 → 0.1867	+0.0036	7.26% → 5.65% Selected 9 → 7	-25.01% → -18.66%
IWM · 09:35	166	0.2321 → 0.2363	+0.0041	1.60% → 7.20% Selected 2 → 9	+48.72% → +8.72%
IWM · 12:00	166	0.2193 → 0.2225	+0.0033	34.40% → 33.60% Selected 43 → 42	-27.91% → -29.31%

Features and sample coverage

Fixed headline counts and topic shares from the earlier experiment. An explicit reference, not a new discovery.

Feature-eligible rows: 1477; missing features: 0; common prediction dates: 165; policy dates: 124; unavailable group / fold fits: 0.

Sensitivity: exclude one calendar month

Excluded month	Selected stress mean · baseline	Selected stress mean · family
2025-07	-27.75%	-22.34%
2025-08	-27.75%	-22.34%
2025-09	-28.37%	-20.70%
2025-10	-28.72%	-20.18%
2025-11	-37.57%	-41.28%
2025-12	-15.71%	-11.28%

Calibration: forecasts and observed frequencies

Counts are ETF / entry observations, not independent trading dates. Sparse bins do not establish calibration.

Matching baseline

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	6	8.07%	50.00%
0.1–0.2	43	16.48%	20.93%
0.2–0.3	351	25.79%	29.34%
0.3–0.4	502	34.04%	33.86%
0.4–0.5	77	43.86%	19.48%
0.5–0.6	11	53.77%	45.45%
0.6–0.7	0	—	—
0.7–0.8	0	—	—
0.8–0.9	0	—	—
0.9–1.0	0	—	—

With this family

Probability bin	N	Mean forecast	Observed positive outcomes
0.0–0.1	21	6.82%	38.10%
0.1–0.2	61	16.57%	27.87%
0.2–0.3	400	26.03%	30.00%
0.3–0.4	415	34.08%	33.73%
0.4–0.5	57	44.13%	22.81%
0.5–0.6	23	54.66%	8.70%
0.6–0.7	10	64.23%	40.00%
0.7–0.8	3	74.36%	33.33%
0.8–0.9	0	—	—
0.9–1.0	0	—	—

Economic assumptions and research status

One same-day ATM call + put bought at asks; ordinary exits use bids of the same contracts at 15:55. Base fee $4 per pair; stress deduction $12 total. Return is net value / paid option premium, not annualized return or return on account capital. A confirmed zero bid is a conservative write-off with no exercise assumed, not an executed sale at zero. Unavailable exits outside the declared write-off policy remain unknown.

Expanding development folds: January–April initial training, then May–June, July–August, September–October and November–December. Penalties and separate ETF/time thresholds use earlier folds with a one-session purge. The first fold has no alert threshold and is excluded from policy returns. Intervals use 1,000 paired bootstrap draws of five-date blocks; they describe development uncertainty, not independent validation. ETF/time rows on one date are dependent and are not summed into an investment portfolio.

H3 event facts are pending a semantically audited full-period feature panel. The LLM pilot is not used here as a validated forecast. Historical PIT limitations remain even if statistical metrics improve.

2026-09-12 19:37:33 UTC
SHA-256: c479bd252eaa7d02372c1a5e7a51b7cb3b5fe57c79e2142e239a418cd1c5106a
Protocol: cd2d96c65dd039aac90ca286e81a28a451351af274a6dcf16da51d76c3bbf6c8
Valuation protocol: 33bdd442f7a600c7daf3bc1e8c670ff78c83cc65fbdb12964612f283117cf185

Does taking profit earlier help?

2025 · TP25

Fixed 15:55 → TP25: one ATM call+put bought at ask. TP25 closes both at the first valid minute with net/debit ≥25%; otherwise the same 15:55 valuation. Base costs $4; stress $12 total, at the same exit.

Development, not a confirmed signal. All six annual TP25 totals remain negative. More positive days do not guarantee profit: an early target can cut the large winners that pay for losses.

1477 identical paired ETF/time cases across 247 dates, not independent days or a combined portfolio. 23 excluded / unavailable cases are not counted as zero.

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 246	-2683 → -3751 Stress: -4651 → -5719	-6.33% → -6.43% Stress: -8.89% → -8.99%	39.02% → 58.54% 96 → 144 / 246
SPY · 12:00 N = 246	-5348 → -7707 Stress: -7316 → -9675	-13.16% → -13.28% Stress: -16.88% → -17.01%	29.67% → 47.97% 73 → 118 / 246
QQQ · 09:35 N = 246	-7433 → -6022 Stress: -9401 → -7990	-10.00% → -7.94% Stress: -12.19% → -10.13%	32.93% → 54.47% 81 → 134 / 246
QQQ · 12:00 N = 246	-8483 → -9765 Stress: -10451 → -11733	-17.69% → -16.14% Stress: -21.05% → -19.50%	24.39% → 43.09% 60 → 106 / 246
IWM · 09:35 N = 247	-5693 → -3360 Stress: -7669 → -5336	-11.91% → -8.89% Stress: -16.37% → -13.35%	34.82% → 56.28% 86 → 139 / 247
IWM · 12:00 N = 246	-5306 → -4806 Stress: -7274 → -6774	-20.13% → -16.01% Stress: -27.18% → -23.06%	29.67% → 50.00% 73 → 123 / 246

Dollars: sum for one pair per session. Percentage: arithmetic mean of each day’s net divided by that day’s paid premium (debit). Different premium sizes can make these comparisons differ. Neither is annual return on required capital; no reinvestment is simulated.

Quarters and January–April

The four quarters partition the year; January–April overlaps Q1/Q2 and must not be added to them. January–April is descriptive development/training data, not forecast validation.

Q1 · 2025

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 60	-1068 → -1703 Stress: -1548 → -2183	-0.58% → -6.34% Stress: -2.82% → -8.58%	38.33% → 58.33% 23 → 35 / 60
SPY · 12:00 N = 60	-1566 → -1845 Stress: -2046 → -2325	-7.52% → -12.08% Stress: -10.78% → -15.35%	33.33% → 50.00% 20 → 30 / 60
QQQ · 09:35 N = 60	-2864 → -1817 Stress: -3344 → -2297	-7.74% → -7.34% Stress: -9.62% → -9.22%	31.67% → 55.00% 19 → 33 / 60
QQQ · 12:00 N = 60	-2466 → -2042 Stress: -2946 → -2522	-14.39% → -13.65% Stress: -17.37% → -16.62%	25.00% → 46.67% 15 → 28 / 60
IWM · 09:35 N = 60	-2709 → -1903 Stress: -3189 → -2383	-19.67% → -15.04% Stress: -23.81% → -19.18%	33.33% → 51.67% 20 → 31 / 60
IWM · 12:00 N = 59	-1269 → -775 Stress: -1741 → -1247	-17.75% → -10.76% Stress: -24.41% → -17.42%	27.12% → 55.93% 16 → 33 / 59

Q2 · 2025

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 61	+11 → -673 Stress: -477 → -1161	-13.44% → -6.96% Stress: -15.46% → -8.98%	37.70% → 57.38% 23 → 35 / 61
SPY · 12:00 N = 62	+953 → -2394 Stress: +457 → -2890	-8.20% → -13.32% Stress: -11.10% → -16.22%	29.03% → 46.77% 18 → 29 / 62
QQQ · 09:35 N = 62	-1550 → -1976 Stress: -2046 → -2472	-17.03% → -11.79% Stress: -18.87% → -13.63%	33.87% → 51.61% 21 → 32 / 62
QQQ · 12:00 N = 62	-477 → -3624 Stress: -973 → -4120	-14.60% → -19.06% Stress: -17.37% → -21.83%	22.58% → 37.10% 14 → 23 / 62
IWM · 09:35 N = 62	-2657 → -889 Stress: -3153 → -1385	-24.71% → -13.14% Stress: -28.81% → -17.24%	30.65% → 53.23% 19 → 33 / 62
IWM · 12:00 N = 62	-1016 → -1874 Stress: -1512 → -2370	-19.32% → -20.95% Stress: -25.74% → -27.38%	25.81% → 41.94% 16 → 26 / 62

Q3 · 2025

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 63	-2152 → -1490 Stress: -2656 → -1994	-14.65% → -10.97% Stress: -18.00% → -14.33%	39.68% → 53.97% 25 → 34 / 63
SPY · 12:00 N = 63	-1894 → -1627 Stress: -2398 → -2131	-15.41% → -12.25% Stress: -20.21% → -17.06%	28.57% → 46.03% 18 → 29 / 63
QQQ · 09:35 N = 63	-4010 → -1527 Stress: -4514 → -2031	-22.09% → -9.26% Stress: -24.93% → -12.10%	30.16% → 53.97% 19 → 34 / 63
QQQ · 12:00 N = 63	-2719 → -2078 Stress: -3223 → -2582	-22.92% → -16.58% Stress: -27.17% → -20.84%	26.98% → 46.03% 17 → 29 / 63
IWM · 09:35 N = 63	-1023 → -385 Stress: -1527 → -889	-9.43% → -6.18% Stress: -14.54% → -11.28%	31.75% → 55.56% 20 → 35 / 63
IWM · 12:00 N = 63	-1308 → -1198 Stress: -1812 → -1702	-18.11% → -16.81% Stress: -26.16% → -24.86%	36.51% → 49.21% 23 → 31 / 63

Q4 · 2025

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 62	+526 → +115 Stress: +30 → -381	3.54% → -1.39% Stress: 0.94% → -3.99%	40.32% → 64.52% 25 → 40 / 62
SPY · 12:00 N = 61	-2841 → -1841 Stress: -3329 → -2329	-21.42% → -15.49% Stress: -25.32% → -19.39%	27.87% → 49.18% 17 → 30 / 61
QQQ · 09:35 N = 61	+991 → -702 Stress: +503 → -1190	7.40% → -3.27% Stress: 5.22% → -5.45%	36.07% → 57.38% 22 → 35 / 61
QQQ · 12:00 N = 61	-2821 → -2021 Stress: -3309 → -2509	-18.68% → -15.15% Stress: -22.09% → -18.57%	22.95% → 42.62% 14 → 26 / 61
IWM · 09:35 N = 62	+696 → -183 Stress: +200 → -679	5.87% → -1.45% Stress: 1.41% → -5.92%	43.55% → 64.52% 27 → 40 / 62
IWM · 12:00 N = 62	-1713 → -959 Stress: -2209 → -1455	-25.27% → -15.25% Stress: -32.31% → -22.29%	29.03% → 53.23% 18 → 33 / 62

January–April · 2025

ETF / entry NY	Total net, $	Mean net/debit	Positive cases
SPY · 09:35 N = 80	+3065 → +390 Stress: +2425 → -250	2.46% → -1.84% Stress: 0.48% → -3.82%	41.25% → 62.50% 33 → 50 / 80
SPY · 12:00 N = 81	+1552 → -1949 Stress: +904 → -2597	-2.54% → -8.84% Stress: -5.39% → -11.70%	34.57% → 53.09% 28 → 43 / 81
QQQ · 09:35 N = 81	+641 → -270 Stress: -7 → -918	-3.50% → -4.40% Stress: -5.20% → -6.10%	35.80% → 58.02% 29 → 47 / 81
QQQ · 12:00 N = 81	+46 → -2670 Stress: -602 → -3318	-8.34% → -11.18% Stress: -10.99% → -13.83%	27.16% → 48.15% 22 → 39 / 81
IWM · 09:35 N = 81	-3053 → -1155 Stress: -3701 → -1803	-18.14% → -10.04% Stress: -21.94% → -13.84%	33.33% → 56.79% 27 → 46 / 81
IWM · 12:00 N = 80	-775 → -1218 Stress: -1415 → -1858	-13.24% → -9.62% Stress: -19.30% → -15.68%	28.75% → 56.25% 23 → 45 / 80

Can the original signals predict TP25 profitability?

Addition to market/cost baseline	Δ Brier 95% CI	Δ stress net/debit, pp per opportunity95% CI
H1 · market path 64 common OOF dates · 44 alert-policy dates	+0.0006[-0.0011; +0.0023]	+0.57[-1.99; +2.45]
H2 · news tempo 165 common OOF dates · 124 alert-policy dates	+0.0005[-0.0018; +0.0028]	+0.61[-0.02; +1.36]
Static headline topics 165 common OOF dates · 124 alert-policy dates	+0.0044[+0.0001; +0.0090]	-0.47[-1.51; +0.34]

Enhanced − matching baseline: negative ΔBrier is better, positive economic Δ is better. Five-session block bootstrap. Skipped opportunities contribute zero to the policy difference; this is not the mean among selected trades. Matching feature cohorts and unchanged expanding folds; forecasts start in May, alerts only after an earlier forecast fold. No confirmed predictive improvement or independent 2026 test.

Quote grid: 435361 / 455100 valid minutes (95.66%); 19739 unavailable. Unobserved touches are not ruled out. Hindsight path highs/lows are not achieved profit. Historical NBBO does not guarantee fills. A negative residual-write-off floor does not prove a real bid-exit loss; abandonment/do-not-exercise is assumed. Historical news availability remains unverified.

Snapshot and verification

2026-09-12 19:53:41 UTC · SHA-256 cba32b05c32596463bd5c1db9554fcb7eaef190eccfcf7b317f242e8d4786c2e

Additional checks · 2025

H4 · H5 · H6 · short

H4: caught 9/47 tail observations, the same number as baseline. H5: all four variants have negative selected-trade stress means.

Short filters: none of the nine checks confirmed an improvement over no filter on its available sample. Q1 and April were training, not out-of-fold tests. Only the separate H6/H2 candidate has a 2026 check above.

H4 · large-move warnings: details

Target: conservative fixed-close long-straddle net/debit ≥100%. Brier covers 2025-05-01 — 2025-12-31 (165 common dates); alarm policy covers 2025-07-01 — 2025-12-31 (124 dates).

Brier: 0.053540 → 0.053513; Δ -0.0000268, 95% date-block interval [-0.0001110; +0.0000938]. Lower is better.

ETF / NY	Baseline: caught / tails	H4: caught / tails	Baseline stress mean	H4 stress mean
SPY · 09:35Brier observations: 166	0 / 9alarms 1 / 125	0 / 9alarms 3 / 125	-44.72%	-14.18%
SPY · 12:00Brier observations: 165	0 / 6alarms 0 / 124	0 / 6alarms 1 / 124	—	-9.32%
QQQ · 09:35Brier observations: 165	0 / 8alarms 0 / 124	0 / 8alarms 0 / 124	—	—
QQQ · 12:00Brier observations: 165	0 / 6alarms 5 / 124	0 / 6alarms 7 / 124	-42.33%	-28.58%
IWM · 09:35Brier observations: 166	9 / 14alarms 83 / 125	9 / 14alarms 83 / 125	-5.20%	-5.20%
IWM · 12:00Brier observations: 166	0 / 4alarms 0 / 125	0 / 4alarms 0 / 125	—	—

The historical 80th-percentile alarm threshold does not guarantee a future 20% alarm rate. Group availability differs; the pooled comparison uses only common dates. Percentages divide net P&L by the option premium paid, not by account capital.

SHA256 4cd9a745740ed5e774d73644306d10cc2d90d48d2686c5552921ca17111287e7

H5 · weekly refitting and last 90 sessions

2025-05-05 — 2025-12-31 · 163 identical common dates for all four variants. Predict the size of net/debit, not the probability of a win; enter only if the expected result remains positive after stress costs. Training uses earlier sessions only, with the preceding session purged.

History / model	Selected stress mean / 95% interval	Selected / dates	Stress per opportunity	Unit-pair P&L sum	MSE
Expanding historyMarket baseline	-33.02%[-44.25%; -17.37%]	104 / 46978 opportunities	-3.51%[-5.54%; -1.57%]	$-12716	0.5727
Expanding historyBaseline + H2 news	-24.13%[-39.69%; -3.35%]	151 / 61978 opportunities	-3.72%[-6.05%; -0.49%]	$-11533	0.5933
Last 90 sessionsMarket baseline	-29.82%[-39.14%; -15.19%]	134 / 69978 opportunities	-4.09%[-5.98%; -1.82%]	$-12419	0.5981
Last 90 sessionsBaseline + H2 news	-31.05%[-39.99%; -15.98%]	190 / 91978 opportunities	-6.03%[-8.27%; -2.53%]	$-16312	0.6343

All return percentages are net/debit on the conservative residual-write-off policy, not account returns or guaranteed market liquidations. Skipped opportunities contribute zero. Dollar sums use one pair per selected ETF/time observation; entry times are alternatives, not a funded portfolio.

Expanding history · ΔMSE +0.0206 [-0.0015; +0.0353]; ΔMAE +0.0064 [-0.0096; +0.0176]. Adding H2 minus baseline; negative error delta is better.

Last 90 sessions · ΔMSE +0.0361 [+0.0065; +0.0590]; ΔMAE +0.0160 [-0.0024; +0.0296]. Adding H2 minus baseline; negative error delta is better.

SHA256 f9c87f18b4660a34a188acbfe6c884e7ede7788689dec94853dc02bd63f83c42

Short · all warning filters and their actual periods (9/9)

Sell at entry bid, close at exit ask using strict fresh quotes; this is not the negative of the long-floor result. Each row is conditional on available quotes and warning history: only 22, 52 or 77 common dates, not a full year. April is outside these evaluated policy slices.

Warning source	Dates / period	Skipped by enhanced filter	Δ$ / opportunity vs no filter	Enhanced − baseline filter	Worst remaining trade
H1 · fixed	222025-09-05 — 2025-12-16	29 / 132	Baseline: -10.92[-28.21; +10.87]Enhanced: -4.64[-12.05; +11.56]	+6.28[-3.17; +23.07]	$-556only this slice
H2 · fixed	522025-07-07 — 2025-12-30	62 / 312	Baseline: -4.04[-12.66; +3.28]Enhanced: -6.54[-18.91; +5.46]	-2.49[-7.83; +3.74]	$-570only this slice
Topics · fixed	522025-07-07 — 2025-12-30	36 / 312	Baseline: -6.01[-14.19; +2.91]Enhanced: -0.20[-7.05; +10.67]	+5.81[-0.97; +13.98]	$-570only this slice
H1 · TP25	222025-09-05 — 2025-12-16	32 / 132	Baseline: -4.60[-15.15; +10.62]Enhanced: -9.10[-24.20; +9.50]	-4.50[-12.26; +1.95]	$-556only this slice
H2 · TP25	522025-07-07 — 2025-12-30	63 / 312	Baseline: -11.81[-23.74; -3.54]Enhanced: -11.70[-28.53; +0.32]	+0.12[-6.56; +6.41]	$-570only this slice
Topics · TP25	522025-07-07 — 2025-12-30	53 / 312	Baseline: -3.05[-12.64; +5.88]Enhanced: -6.31[-14.24; +1.14]	-3.26[-9.03; +2.98]	$-570only this slice
H4 · tail	522025-07-07 — 2025-12-30	45 / 312	Baseline: -3.41[-8.15; +1.01]Enhanced: -4.54[-8.75; -0.64]	-1.12[-3.14; +0.00]	$-570only this slice
H5 · expanding	772025-05-05 — 2025-12-30	86 / 462	Baseline: -17.20[-30.00; -5.18]Enhanced: -15.61[-26.54; -0.55]	+1.59[-4.12; +12.93]	$-676only this slice
H5 · 90 sessions	772025-05-05 — 2025-12-30	105 / 462	Baseline: -13.44[-26.55; -0.01]Enhanced: -16.28[-26.19; -0.22]	-2.84[-6.72; +8.33]	$-676only this slice

Intervals use paired five-session date blocks. Positive dollar delta is better. These incomplete cohorts differ and must not be ranked as equivalent full-year strategies; the worst shown trade is not an annual maximum loss. All checks are 2025 development, with no verified historical news availability.

SHA256 9f214250b5cc38f8922e89780637369aaac1bd43a1601e915770f808a7aec5c3

SHA256 296a89669c952435c4fc33021651ea13a274026967d3960387dacd835312c69c

H6 · future realised volatility / premium-cost proxy: not a profit forecast

Full source: 1480 observations, 326 positive labels. Out-of-fold scores: 166 dates / 996 observations, 2025-05-01 — 2025-12-31. Alarm-policy sample: 750 / 169 positive cases, 2025-07-01 — 2025-12-31.

ΔBrier H2 − baseline: -0.001326 [-0.004157; +0.000645]. The interval includes zero: improvement is not established. Precision +22.39% → +31.39%; recall +8.88% → +25.44%; alarm coverage +8.93% → +18.27%.

Future minute-return RV exceeds a Gaussian ATM premium-cost proxy. This proxy is not measured IV, and high realised volatility is not a profitable straddle. There is no option P&L claim here.

ETF / NY	Baseline precision / recall	H2 precision / recall	Baseline alarms / rows	H2 alarms / rows
SPY · 09:35	— / +0.00%	+30.00% / +9.09%	0 / 125	10 / 125
SPY · 12:00	+27.27% / +12.00%	+26.67% / +16.00%	11 / 125	15 / 125
QQQ · 09:35	+0.00% / +0.00%	+35.29% / +18.18%	1 / 125	17 / 125
QQQ · 12:00	+14.29% / +9.52%	+20.00% / +23.81%	14 / 125	25 / 125
IWM · 09:35	+26.47% / +24.32%	+39.58% / +51.35%	34 / 125	48 / 125
IWM · 12:00	+14.29% / +5.00%	+27.27% / +30.00%	7 / 125	22 / 125

More true positives came with more alarms; this is not a 70–80% forecast. The historical quantile threshold is not a fixed future alarm rate. Q1/April were training; all results remain development-only.

SHA256 4a205838091d462c90396c79cbbb8137a200e2ecaa70f1889944c1acc24ba5a8

News annotation quality has only a small single-reviewer audit; agreement with that reviewer is not forecast accuracy.

2025: news before the move

SPY · QQQ · IWM

Do headlines known before a decision time relate to subsequent volatility, and do they add information beyond the market itself? This first study uses a fixed dictionary of headline topics, not LLM scores.

Exploratory historical study, not proof of a tradable forecast. Historical text was downloaded later and its original availability is unverified. Correlation, including partial correlation, does not establish causality. The 600-article LLM pilot below is a separate experiment.

Source corpus: 566378 URLs; 538196 pass metadata admission; 445532 appear in at least one of the 500 source windows, including early-close days. These are not independent market observations.

Trading dates: 247ETF / cutoff observations: 1482Excluded early-close observations: 18

Does adding news improve the held-out test?

All six ETF / time combinations are shown, without selecting only winners. Models predict log(range % + 0.01) on July–December dates. Smaller MAE is better; positive improvement means adding news helped. This is an error metric, not the percentage of correctly predicted days.

ETF / NY	N	Market MAE	News MAE	Combined MAE	Improvement	Δ MAE · 95% CI
QQQ · 09:25	125	0.327	0.499	0.368	-12.5%	+0.041 [+0.019; +0.061]
QQQ · 12:00	125	0.421	0.657	0.452	-7.4%	+0.031 [+0.009; +0.052]
SPY · 09:25	125	0.338	0.509	0.379	-12.3%	+0.041 [+0.020; +0.064]
SPY · 12:00	125	0.418	0.634	0.447	-7.1%	+0.030 [+0.011; +0.047]
IWM · 09:25	125	0.322	0.345	0.310	+3.5%	-0.011 [-0.028; +0.006]
IWM · 12:00	125	0.377	0.453	0.385	-2.1%	+0.008 [-0.006; +0.025]

Δ MAE = combined − market: negative is better. Intervals use paired 5-session blocks and 1,000 bootstrap samples; crossing zero means the sign is uncertain. ETF observations on one date are correlated, not independent tests.

Pooled date-level improvement: -6.4% · 125 dates. Δ MAE +0.023 [+0.009; +0.037].

Inspect each feature

Headline feature	N	Spearman ρ	Partial ρ
All headlines: log(1 + count)	125	+0.338	+0.169
Macro / rates	125	+0.027	+0.070
Trade policy / tariffs	125	-0.171	-0.162
Geopolitics	125	-0.057	-0.012
Earnings / guidance	125	+0.161	+0.045
Uncertainty / risk	125	+0.054	+0.066

Topic features are shares of unique normalized headlines in the prior 24 hours; topics may overlap. Positive ρ means larger subsequent volatility, negative means smaller. Partial ρ controls for prior market volatility, the premarket gap and weekday; noon also controls for the completed morning. These are descriptive associations, not probabilities. The upper model comparison always targets range, even when this table selects realized volatility.

Daily data for this selection · 125

Date	Range %	RV, %	Headlines	Macro %	Trade %	Geopolitics %	Earnings %	Risk %
2025-07-01	0.569	0.447	1477	2.7	2.2	1.0	5.5	2.4
2025-07-02	0.204	0.258	1451	2.1	2.3	0.8	7.0	2.6
2025-07-07	0.788	0.579	947	1.1	4.2	0.8	12.6	2.7
2025-07-08	0.396	0.386	1352	1.3	6.4	1.0	6.7	2.9
2025-07-09	0.513	0.303	1426	2.7	5.5	1.1	5.3	3.4
2025-07-10	0.370	0.266	1571	1.3	5.3	1.1	5.7	2.6
2025-07-11	0.416	0.339	1423	2.0	5.9	1.3	8.2	3.0
2025-07-14	0.285	0.226	1070	2.5	3.9	0.7	6.7	3.8
2025-07-15	0.553	0.370	1609	3.3	5.2	0.9	6.0	3.8
2025-07-16	0.638	0.378	1679	3.7	5.4	0.9	10.5	3.4
2025-07-17	0.196	0.225	1741	2.8	4.0	0.6	9.9	3.2
2025-07-18	0.333	0.307	1664	2.9	2.9	0.7	12.7	3.2
2025-07-21	0.400	0.216	1222	2.3	2.9	0.7	11.8	3.7
2025-07-22	0.421	0.332	1687	2.3	3.4	0.7	11.7	3.1
2025-07-23	0.557	0.332	1882	1.3	4.9	0.8	16.9	3.1
2025-07-24	0.348	0.275	1901	1.7	3.7	0.7	10.4	2.3
2025-07-25	0.289	0.197	1860	2.0	3.9	0.4	16.6	3.2
2025-07-28	0.283	0.222	1301	1.5	6.0	0.4	13.2	2.2
2025-07-29	0.433	0.309	1863	1.9	6.0	0.8	18.0	3.0
2025-07-30	0.979	0.575	2192	2.1	4.9	0.7	22.1	2.1
2025-07-31	1.323	0.650	2271	4.0	5.5	0.4	18.3	2.2
2025-08-01	1.106	0.648	2226	2.3	6.4	0.6	21.2	2.2
2025-08-04	0.289	0.285	1389	1.3	3.0	0.5	20.8	2.8
2025-08-05	0.586	0.482	1922	2.5	2.9	0.6	20.2	2.5
2025-08-06	0.421	0.323	2048	1.2	3.8	0.6	25.4	1.9
2025-08-07	0.780	0.643	2139	1.7	5.6	0.8	20.0	2.2
2025-08-08	0.286	0.300	2182	2.5	3.7	0.5	27.3	1.7
2025-08-11	0.908	0.376	1280	3.0	3.4	0.9	14.0	4.1
2025-08-12	0.426	0.225	2009	4.2	4.9	0.5	16.1	2.5
2025-08-13	0.293	0.351	2868	2.4	3.7	0.4	32.4	2.3
2025-08-14	0.478	0.379	2178	2.7	2.2	0.4	26.2	2.1
2025-08-15	0.352	0.345	1892	3.2	2.3	0.7	25.6	2.9
2025-08-18	0.387	0.287	1227	2.0	1.5	1.1	10.3	3.7
2025-08-19	0.640	0.430	1524	1.6	1.6	1.0	8.4	2.1
2025-08-20	0.859	0.602	1757	2.7	2.4	0.5	13.2	2.8
2025-08-21	0.574	0.492	1850	3.6	3.1	0.4	12.5	2.5
2025-08-22	0.496	0.361	1694	4.1	2.1	0.7	12.7	3.4
2025-08-25	0.534	0.271	1240	3.3	1.9	0.9	9.4	4.2
2025-08-26	0.503	0.320	1798	4.9	2.0	0.5	9.0	3.8
2025-08-27	0.480	0.314	1860	4.5	2.2	0.6	13.3	2.4
2025-08-28	0.506	0.269	1966	2.1	2.1	0.7	16.1	3.2
2025-08-29	0.405	0.416	1886	3.6	3.0	1.0	16.0	2.8
2025-09-02	1.094	0.498	1446	2.6	1.9	1.1	9.3	3.8
2025-09-03	0.618	0.473	1717	3.3	2.3	0.5	8.6	2.2
2025-09-04	0.713	0.280	1804	2.8	2.4	0.4	11.0	2.8
2025-09-05	0.458	0.562	1916	4.1	2.7	0.5	7.8	2.2
2025-09-08	0.470	0.341	1440	4.4	1.9	0.6	5.8	2.8
2025-09-09	0.443	0.261	2314	2.9	1.4	0.5	3.8	2.9
2025-09-10	0.680	0.380	1873	3.5	1.2	0.5	6.0	2.0
2025-09-11	0.205	0.242	2114	4.5	1.2	0.4	6.7	2.2
2025-09-12	0.357	0.213	1902	3.4	1.7	0.6	5.7	2.5
2025-09-15	0.288	0.221	1598	4.0	1.3	0.6	5.8	1.9
2025-09-16	0.267	0.193	1722	5.3	1.9	0.9	4.8	2.9
2025-09-17	1.255	0.895	1893	5.4	1.0	0.7	4.9	2.3
2025-09-18	0.402	0.312	1866	7.7	1.4	0.6	4.5	2.7
2025-09-19	0.455	0.270	1661	4.2	1.7	0.7	4.1	2.5
2025-09-22	0.463	0.240	1354	3.1	1.3	0.7	3.8	3.8
2025-09-23	0.702	0.402	1771	3.1	1.9	0.7	3.7	3.0
2025-09-24	0.523	0.341	2087	2.7	1.2	0.6	4.8	3.2
2025-09-25	0.965	0.510	1724	2.9	1.6	0.6	3.9	2.3
2025-09-26	0.680	0.370	1788	3.5	3.5	0.6	4.1	2.5
2025-09-29	0.359	0.314	1390	2.4	1.7	0.6	4.7	3.2
2025-09-30	0.732	0.370	1766	2.9	1.8	0.5	3.6	2.7
2025-10-01	0.429	0.262	1761	2.3	1.8	0.7	3.4	2.5
2025-10-02	0.360	0.260	2012	2.3	1.4	0.7	3.2	2.0
2025-10-03	0.928	0.487	1841	2.2	1.4	0.5	5.4	2.8
2025-10-06	0.286	0.236	1472	1.0	1.3	0.5	6.2	3.3
2025-10-07	0.401	0.449	1731	1.4	1.8	0.8	5.2	3.4
2025-10-08	0.446	0.222	1799	2.2	1.7	0.3	5.1	2.7
2025-10-09	0.538	0.383	1936	1.9	1.6	0.8	5.7	3.9
2025-10-10	1.862	1.013	1829	2.4	2.0	0.9	6.0	3.4
2025-10-13	0.373	0.422	1318	1.7	3.3	1.2	7.1	4.8
2025-10-14	1.055	0.713	1829	1.7	3.4	1.0	5.8	4.3
2025-10-15	1.306	0.803	1919	2.4	2.9	0.9	7.5	4.0
2025-10-16	1.402	1.098	1932	2.5	2.2	1.1	8.1	3.6
2025-10-17	0.897	0.711	1919	2.1	2.6	0.7	8.1	4.3
2025-10-20	0.250	0.237	1153	2.2	2.4	1.0	8.1	4.6
2025-10-21	0.384	0.316	1650	2.1	2.2	0.7	11.2	2.4
2025-10-22	1.208	0.921	2160	1.5	1.5	0.6	15.2	2.3
2025-10-23	0.399	0.321	2350	1.1	1.8	0.5	14.6	2.9
2025-10-24	0.271	0.226	2297	2.4	2.6	0.7	19.5	2.4
2025-10-27	0.358	0.189	1371	3.3	3.1	0.6	12.9	3.1
2025-10-28	0.655	0.344	2372	2.6	2.3	0.6	13.7	2.2
2025-10-29	0.914	0.895	2671	2.4	1.3	0.3	17.8	2.0
2025-10-30	0.876	0.517	2696	3.7	2.2	0.5	14.3	2.3
2025-10-31	0.880	0.606	3217	1.3	1.0	0.4	33.9	2.1
2025-11-03	0.377	0.339	1791	1.5	1.7	0.5	16.4	2.8
2025-11-04	0.831	0.582	2389	1.5	1.0	0.4	21.1	2.6
2025-11-05	0.567	0.340	3006	0.8	1.5	0.4	32.6	1.8
2025-11-06	0.954	0.742	3006	1.1	1.0	0.2	26.8	2.0
2025-11-07	1.881	0.777	2931	1.2	1.0	0.3	27.9	2.8
2025-11-10	0.981	0.390	1592	1.5	2.1	0.8	12.7	2.8
2025-11-11	0.733	0.460	2179	0.9	1.7	0.4	21.5	2.3
2025-11-12	0.507	0.507	2152	1.3	1.5	0.3	19.6	2.6
2025-11-13	0.999	0.857	2364	2.0	1.1	0.5	24.4	2.7
2025-11-14	0.877	0.792	2443	1.9	1.1	0.6	22.4	2.5
2025-11-17	1.844	0.771	1469	1.8	2.0	0.5	11.0	4.6
2025-11-18	1.286	1.060	1965	1.2	1.0	0.7	12.7	4.1
2025-11-19	1.293	0.876	1951	1.8	0.9	0.5	11.7	3.7
2025-11-20	2.209	1.640	2105	3.0	1.1	0.6	13.4	2.8
2025-11-21	1.609	1.048	1852	2.5	1.5	1.0	10.7	3.8
2025-11-24	0.663	0.424	1394	2.8	0.6	0.6	6.4	4.2
2025-11-25	0.969	0.427	1812	2.6	0.9	0.5	8.0	3.6
2025-11-26	0.343	0.425	1807	2.7	0.7	0.5	10.6	2.7
2025-12-01	0.547	0.405	1381	2.6	0.7	0.5	3.5	3.6
2025-12-02	0.588	0.456	1901	1.5	1.7	1.0	6.1	3.2
2025-12-03	0.505	0.293	1945	1.9	1.3	0.6	6.4	2.9
2025-12-04	0.631	0.430	2059	1.8	0.9	0.8	5.0	4.3
2025-12-05	0.363	0.352	2188	1.6	0.8	0.7	7.6	3.1
2025-12-08	0.510	0.406	1584	2.9	1.1	0.5	5.1	3.6
2025-12-09	0.198	0.247	1941	2.5	1.2	0.9	4.7	3.5
2025-12-10	1.061	0.759	2070	3.9	0.5	0.5	5.8	3.4
2025-12-11	0.886	0.415	2063	5.9	0.8	0.3	6.3	3.2
2025-12-12	0.889	0.605	1905	3.1	0.8	0.7	6.7	3.1
2025-12-15	0.421	0.521	1419	2.7	0.8	0.4	3.7	3.5
2025-12-16	1.081	0.596	1825	2.6	0.8	0.5	4.8	3.6
2025-12-17	0.806	0.609	1932	2.8	0.4	0.6	4.8	3.1
2025-12-18	0.848	0.675	2145	3.1	0.7	0.7	5.8	2.7
2025-12-19	0.361	0.280	1918	3.1	0.8	0.7	4.5	2.8
2025-12-22	0.203	0.193	1327	1.8	1.3	0.5	4.4	3.3
2025-12-23	0.277	0.153	1690	2.1	0.8	0.4	4.0	2.7
2025-12-26	0.187	0.183	719	0.8	1.1	0.8	3.9	2.8
2025-12-29	0.478	0.236	996	1.8	0.4	0.9	2.6	4.2
2025-12-30	0.293	0.194	1156	1.0	0.6	0.9	2.9	3.2
2025-12-31	0.736	0.347	1221	1.7	0.7	0.7	3.2	2.7

Method and reproducibility

Massive unadjusted minute bars; America/New_York calendar. At 09:25 the target starts at the 09:30 open, at noon it starts at 12:00. Features use only completed prior minutes. Early-close sessions are excluded from this comparison. Missing candles are not filled. News availability uses the later publication/update time plus 5 minutes.

Development: January–April; regularization chosen on May–June; refit before the frozen July–December test, with a one-session boundary purge. The dictionary is fixed before joining market outcomes. This study does not use or assess the LLM pilot scores and does not establish trading profitability.

2026-09-12 18:33:03 UTC
SHA-256: 646772db931958fb0bac8b0c989eaf5d5663a60015832b93196a0157314d8c9c
Protocol: 4a95fc1a44618ec535d2789b49feeba74e7903d778cdb2b28fb1fd66fb349097

Separate LLM extraction pilot below. A valid model response is not a correct prediction. The LLM scores have not been validated as forecasts or used in the headline correlation study above.

Pilot: processed / selected600 / 600Including explicitly skipped/error cases

Both model stages valid586Format and source binding only

Nasdaq relevance: yes / uncertain268 / 33First model assessment

Forecast accuracy—Not tested yet

Article collection

Yahoo Finance → separate research archive

Archive year	Found URLs	Ready	Partial	Errors / unavailable
2022 2022-01-01 — 2022-12-31	1 507	1 507	0	0
2023 2023-01-01 — 2023-12-31	164 417	164 263	13	141
2024 2024-01-01 — 2024-12-31	482 008	461 939	19 877	192
2025 2025-01-01 — 2025-12-31	566 484	538 474	27 922	88
2026 2026-01-01 — 2026-10-16	583 864	562 670	21 074	120

Snapshot: 2026-10-17 07:21:50 UTC. Unique URLs discovered on archive dates in that year, not unique news events. “Partial” commonly means a source-restricted article. “Ready” does not certify complete original text or historical availability.

Current collector: collected_with_gaps · status outdatedNew ready in this process run: 2 616Queued / being processed: 0

Collector status age: 60 955 s. These run counters reset on process restart and do not represent the full corpus or a full archive year.

Local-model pilot

Completed with errors

Qwen3.6-35B-A3B: broad Nasdaq relevance filter. Qwen3.6-27B: claims, importance and potential volatility impact. Thinking off; temperature 0. In this diagnostic pilot the second model also reads filter rejections, to compare decisions and flag possible misses for review.

100.0% · 600/600Filter / scorer errors: 0 / 14Source errors / excluded: 0 / 0

Updated: 2026-09-12 17:56:01 UTC. Strict point-in-time eligible: 0. All other results are exploratory only.

Publication month	Sample	Processed	Relevant	Not relevant	Uncertain
2025-01	150	150	69	68	13
2025-02	150	150	53	91	6
2025-03	150	150	69	73	8
2025-04	150	150	77	67	6

A bounded extraction sample, not the whole monthly corpus and not a representative estimate of news frequency.

Second model, Nasdaq relevance: yes / no / uncertain: 4 / 139 / 443. Filter says no, scorer says yes or uncertain: 242 / 586. These need review; they are not proven filter errors.

Importance distribution

Ordinal model score 0–10, not a probability of a market move.

200

552

203

4775

010

The same score 5/10 appears in 477 / 586 valid scorer responses. We are checking whether the scores distinguish events; this is not evidence of quality.

Potential volatility impact

Low	144
Medium	0
High	89
Uncertain	353

Text assessment, not measured future volatility. Market outcomes have not been joined to these scores.

Examples for checking the extraction

At most 30 · highest importance first

Model claims are not verified facts. A source excerpt is bound to the saved text by code; this does not prove that the model interpretation is correct. Selection for display is not a signal backtest.

Article / evidence	Filter	Importance	Volatility
Liang Wenfeng 2025-04-16 10:34:10 UTC · Earnings · Describes moves already made Model claims and source excerpt Nvidia lost $600 billion in market value, the largest drop for a single company in U.S. stock-market history. DeepSeek's low-cost AI helped add billions to the value of tech stocks in China and Hong Kong. DeepSeek surpassed ChatGPT to become the No. 1 free app on Apple's U.S. app store by February. chips produced by Nvidia, which itself quickly lost $600 billion in market value, the largest drop for a single company in U.S. stock-market history. Meanwhile,	Relevant	8/10	High
The Interesting Silver Lining Behind This Painful Tariff Chaos 2025-04-04 16:29:10 UTC · Trade policy · Describes moves already made Model claims and source excerpt U.S. President Trump announced a 10% minimum universal tariff and country-specific tariffs exceeding 50%, prompting China to declare additional 34% tariffs on the U.S. Major U.S. stock indices declined significantly, with the Dow Jones sliding over 3%, the Nasdaq Composite dropping more than 4%, and the S&P 500 sinking more than 4%. The 10-year U.S. Treasury yield collapsed from recent highs near 4.8% to approximately 3.93%, a drop of over 80 basis points, which the article attributes to policy design to lower borrowing costs. "Liberation Day" was a mess… The new levies that U.S. President Trump announced were far more aggressive than anyone had expected. A 10% minimum universal	Relevant	8/10	High
The pursuit of ‘lean’ operations has left companies mercilessly exposed to the tariffs chaos 2025-04-09 15:33:22 UTC · Trade policy · Commentary Model claims and source excerpt President Donald Trump imposed 104% tariffs on China, 10% on general imports, 20% on EU goods, and 46% on Vietnam goods. China announced it will raise tariffs on American goods to 84% starting the day after the US announcement. The White House announced a 90-day pause on some higher tariffs, excluding China. The global trade ecosystem has been overturned. President Donald Trump imposed 104% tariffs on China on Wednesday, a week after levying a host of tariffs	Relevant	8/10	High
Trump Pauses Tariffs for 75 Countries Negotiating With US – but Hits China With 125% – as Market Rebound Surges 2025-04-09 18:16:10 UTC · Trade policy · Describes moves already made Model claims and source excerpt Trump announced a 90-day pause on tariffs for 75 negotiating countries, reducing penalties to 10%, while simultaneously raising tariffs on China to 125%. The Dow Jones Industrial Average surpassed 40,000 points, the S&P 500 reached a high of 5,393, and the Nasdaq rose nearly 10% following the announcement. Media and tech stocks surged, with Comcast, Apple, and Netflix shares increasing more than 6% and WB Discovery up 17.69%. Donald Trump on Wednesday announced a 90-day "pause" on tariffs for countries who are actively negotiating new trade deals with the United States, reducing penalties	Relevant	8/10	High
US dollar is latest to feel impact of Trump tariff moves 2025-04-12 11:01:53 UTC · Trade policy · Describes moves already made Model claims and source excerpt A broad trade-weighted measure of the dollar fell to its lowest level since April 2022, down 10% since Trump's inauguration. The 10-year Treasury yield broke past 4.50% and the 30-year yield briefly hit 5% amid widespread selling of U.S. Treasuries. Gold surged to a new all-time high above $3,200 per ounce while the University of Michigan's preliminary consumer sentiment index for April plunged to its lowest level since mid-2022. Volatility sparked by President Donald Trump's aggressive trade tariffs has extended beyond Wall Street, rattling the Treasury market and the U.S. dollar over the week.	Relevant	8/10	High
Trump’s Powell threats sink stocks in rocky start to trading week 2025-04-21 20:07:49 UTC · Geopolitics · Describes moves already made Model claims and source excerpt U.S. stock indexes fell on Monday amid concerns over tariffs and threats to Federal Reserve independence, with the Dow down 2.48%, S&P 500 down 2.36%, and Nasdaq down 2.55%. President Trump criticized Fed Chair Jerome Powell on social media, calling for lower interest rates and falsely claiming there is virtually no inflation, while White House officials study the possibility of removing Powell. The dollar fell to a three-year low and gold soared to record highs as investors worry about a potential U.S. recession caused by proposed tariffs, with Apollo Global Management estimating a 90% chance of recession in 2025. U.S. stock indexes fell Monday, kicking off a rough start to the week amid growing concerns over tariffs and the future of the U.S. Federal	Relevant	8/10	High
What to Make of the Weakening Economy and Trade War 2025-03-05 17:27:34 UTC · Macro · Commentary Model claims and source excerpt The Atlanta Fed's GDPNow model estimates Q1 2025 real GDP growth at -1.5%, down from 2.3% on Feb 19, with net exports contribution falling to -3.70 percentage points. Trump has enforced 25% tariffs on Mexico and Canada and additional tariffs on China, raising the average U.S. tariff rate from 2.3% to 11.5%, the highest since WWII. The S&P 500 has fallen about 3% in the past five days, testing its 100-day moving average and approaching its 200-day moving average. Right now, the U.S. economy is slowing rapidly. According to the Atlanta Federal Reserve's GDPNow model, "the estimate for real GDP growth (seasonally adjusted annual	Relevant	8/10	High
Five market signals flashing red amid DeepSeek sell-off 2025-01-31 06:41:03 UTC · Macro · Describes moves already made Model claims and source excerpt Nvidia suffered the biggest one-day loss of value on record amid a tech sell-off driven by the popularity of China's DeepSeek AI model. Gross leverage among hedge funds trading U.S. stocks hit their highest level since 2010 in January. The equity risk premium dropped below zero in December for the first time since 2002. LONDON (Reuters) - In a week when AI chipmaker Nvidia (NVDA) suffered the biggest one-day loss of value on record and the Federal Reserve said	Relevant	7/10	High
Black Diamond Group (TSE:BDI) Has Affirmed Its Dividend Of CA$0.035 2025-03-13 10:17:42 UTC · Earnings · Describes moves already made Model claims and source excerpt Black Diamond Group affirmed a dividend of CA$0.035 to be paid on April 15. Earnings per share are forecast to rise by 63.8% over the next year. The annual dividend declined approximately 84% from CA$0.90 in 2015 to CA$0.14. Black Diamond Group (TSE:BDI) Has Affirmed Its Dividend Of CA$0.035	Not relevant	5/10	Uncertain
Accelerating Action: Why Gender Equality Can’t Wait 2025-03-07 12:37:07 UTC · Other · Commentary Model claims and source excerpt The number of Top 100 accountancy firms with female leaders has plummeted by 20% in a year, leaving just 12 women at the helm. The gender pay gap across European financial services boardrooms has widened from 31% to 36% between 2019 and 2023. Female founder teams only receive 2.8% of equity investment in the UK, while the 2023 Rose Review estimated £250 billion could be added to the UK economy if women matched men in entrepreneurship rates. gender equality in recent years. The Hampton and Davies Reviews have driven progress in FTSE company leadership, with women now holding 43% of board positions	Not relevant	5/10	Uncertain
Riskified Ltd. (RSKD) and IXOPAY Partner to Revolutionize E-Commerce Security with AI-Driven Fraud Prevention 2025-01-10 16:37:41 UTC · Other · Describes moves already made Model claims and source excerpt Riskified Ltd. and IXOPAY formed a strategic partnership to improve payment security and reduce fraud in e-commerce. The partnership combines IXOPAY's payment orchestration with Riskified's AI-driven fraud detection to improve transaction efficiency and minimize chargebacks. Super.com is cited as a company already benefiting from the partnership to secure transactions while maintaining PCI compliance. risk management platform using machine learning to protect merchants from fraud, policy abuse, and chargebacks. The company's AI-powered solutions include Policy Protect, Account S	Relevant	5/10	Uncertain
Best Stock to Buy Right Now: FuboTV vs. Netflix 2025-04-24 12:33:00 UTC · Earnings · Reported fact Model claims and source excerpt FuboTV reported 2024 revenue of $1.62 billion, a 19% year-over-year increase, and ended the year with nearly 1.7 million subscribers in North America. Netflix reported first-quarter revenue of $10.5 billion, a 13% year-over-year increase, and net income of $2.9 billion. FuboTV entered an agreement with Disney where Disney receives 70% ownership in Fubo in exchange for Fubo management running Hulu+ Live TV and adding ESPN content. in streaming live sporting events to differentiate itself from other video platforms. The Disney partnership puts Fubo management in charge of Disney's Hulu+ Live TV	Not relevant	5/10	Low
Krispy Kreme (DNUT) Stock Trades Down, Here Is Why 2025-02-25 19:07:14 UTC · Earnings · Describes moves already made Model claims and source excerpt Krispy Kreme shares fell 26.1% in pre-market trading after missing Q4 2024 sales, operating profit, and earnings targets, with revenue down 10.4% year-over-year. The company cited a cybersecurity incident costing an estimated $11 million in lost revenue and the divestiture of Insomnia Cookies as factors in the sales decline. Full-year guidance for revenue, EPS, and EBITDA fell short of Wall Street expectations, leading to a 22% drop in the stock price by close. Krispy Kreme (DNUT) Stock Trades Down, Here Is Why What Happened? Shares of doughnut chain Krispy Kreme (NASDAQ:DNUT) fell 26.1% in the pre-market session after	Not relevant	5/10	High
Napier Port Holdings (NZSE:NPH) Might Be Having Difficulty Using Its Capital Effectively 2025-03-17 20:02:32 UTC · Earnings · Describes moves already made Model claims and source excerpt Napier Port Holdings reported a Return on Capital Employed (ROCE) of 6.4% for the trailing twelve months to September 2024, calculated from NZ$35m EBIT and capital employed of NZ$557m. The company's ROCE has decreased from 8.5% five years ago to 6.4% currently, despite increases in both revenue and assets employed. Shares of Napier Port Holdings delivered a 22% gain to shareholders over the last five years. What Is Return On Capital Employed (ROCE)? For those who don't know, ROCE is a measure of a company's yearly pre-tax profit (its return), relative	Not relevant	5/10	Uncertain
Australian Vintage Ltd (ASX:AVG) (H1 2025) Earnings Call Highlights: Strategic Moves and ... 2025-02-24 03:00:35 UTC · Earnings · Describes moves already made Model claims and source excerpt Australian Vintage Ltd is currently not cash positive and faces concerns about debt levels, though management states debt is consistent with the past five years. The company implemented a $9 million cost-out program last year and anticipates taking out another $6 to $10 million in costs from the next vintage. Australian Vintage Ltd is focusing on industry consolidation and growing the business rather than liquidating assets, despite a significant decline in share price over the last 12 months. company is currently not cash positive, with a focus on achieving cash flow neutrality and positive cash flow in the future. There are concerns about	Not relevant	5/10	Uncertain
Consumer sentiment is plunging. Wall Street sees a tired US shopper as a major risk to stocks. 2025-03-26 02:26:36 UTC · Macro · Describes moves already made Model claims and source excerpt The Consumer Confidence Index dropped to 92.9 in March, its lowest level since 2021, with the Expectations Index falling to 65.2. UBS Global Research forecasts the S&P 500 could drop 8% to 5,300 and 12-month forward earnings growth expectations could fall to 6%. Consumer discretionary stocks in the S&P 500 are down 9% year-to-date, while 70% of companies reporting Q1 earnings issued negative outlooks. Consumer Confidence clocked in at 92.9 in March, its lowest level since 2021, according to the Conference Board's latest survey. Consumer confidence dropped to its	Relevant	5/10	Uncertain
Get a Slice of International Diversification With This Vanguard ETF 2025-02-05 10:30:00 UTC · Macro · Describes moves already made Model claims and source excerpt President Trump is attempting to reset U.S. trade relationships with major partners including China, the European Union, the U.K., Mexico, and Canada. Goldman Sachs and other major investment firms forecast around a 10% total return for the S&P 500 in 2025. The Vanguard Total International Stock ETF (VXUS) has a dividend yield of about 3.4% and trades at 15.5 times earnings, compared to the S&P 500's 1.2% yield and 28.7 earnings multiple. Global stock markets have gotten off to a healthy start in 2025. Late last year, Goldman Sachs, along with several other major investment firms, forecast	Relevant	5/10	Uncertain
Allot Ltd. (ALLT) Helps Rakuten Mobile with AI-Powered Network Protection 2025-03-30 21:27:46 UTC · Earnings · Describes moves already made Model claims and source excerpt Allot Ltd. announced that Rakuten Mobile deployed Allot Network Protection and Deep Network Intelligence solutions during its transition to the cloud-native Rakuten Symphony Platform. The Allot Smart NetProtect solution utilizes ML and AI to detect anomalies in network traffic before they impact service quality. Allot Ltd. is described as a close technology partner supporting Rakuten Mobile in the live cloud-native deployment of fully containerized solutions. announced that Rakuten Mobile deployed Allot Network Protection and Deep Network Intelligence solutions as it transitions from a Network Functions Virtualization architecture to it	Relevant	5/10	Low
Analyst Points Out AI Opportunities for Advanced Micro Devices (AMD) 2025-01-11 20:37:32 UTC · Earnings · Describes moves already made Model claims and source excerpt Anastasia Amoroso stated that 2025 will focus on AI software, data centers, and power, with the AI rally broadening out from semiconductors. David Nelson suggested Advanced Micro Devices, Inc. (NASDAQ:AMD) may play a role in the inference side of AI as companies broaden usage beyond training. Rogue Funds sold Advanced Micro Devices, Inc. (NASDAQ:AMD) puts for a profit after volatility spiked due to the Japan Carry Trade. We recently published a list of 10 Latest AI News You Shouldn't Miss. In this article, we are going to take a look at where	Relevant	5/10	Low
With EPS Growth And More, Smartgroup (ASX:SIQ) Makes An Interesting Case 2025-01-07 05:37:38 UTC · Earnings · Describes moves already made Model claims and source excerpt Smartgroup grew EPS by 10.0% per year over the last three years and revenue by 24% to AU$283m while maintaining stable EBIT margins. Insiders invested AU$849k more than they sold, with Independent Non-Executive Chairman John Prendiville purchasing AU$584k of shares at approximately AU$8.34 per share. CEO Scott Wharton received AU$1.2m in compensation for the year ending December 2023, which is below the median of AU$1.7m for similarly sized companies. trends in share price. That makes EPS growth an attractive quality for any company. Over the last three years, Smartgroup has grown EPS by 10.0%	Not relevant	5/10	Uncertain
Trade Desk’s big bet on AI in advertising hits rollout snags 2025-02-18 10:32:00 UTC · Earnings · Reported fact Model claims and source excerpt The Trade Desk reported Q4 revenue of $741 million, missing internal expectations of at least $756 million for the first time in eight years. The rollout of Kokai, an AI-powered platform overhaul, has been slower than anticipated, with the company aiming to transition all clients by the end of 2025. The Trade Desk's stock price tumbled following the earnings miss, attributed by CEO Jeff Green to execution missteps and a deliberate pace in Kokai's rollout. This story was originally published on Marketing Dive. To receive daily news and insights, subscribe to our free daily Marketing Dive newsletter. Dive Brief: The	Uncertain	5/10	High
Orange Wholesale to redefine mobile connectivity with 5G core network-as-a-service 2025-03-12 11:17:01 UTC · Technology · Reported fact Model claims and source excerpt Orange Wholesale unveiled a fully hosted and managed 5G Core Network-as-a-Service solution at MWC 2025, enabling operators to deploy 5G SA infrastructure without maintaining their own core network. The solution offers a pay-as-you-grow model with end-to-end support, leveraging 51 Telco Cloud Points of Presence and roaming agreements with 130 operators to reduce capital expenditure and financial risk. The launch addresses slow 5G Standalone adoption, with only about 125 operators across 55 countries having deployed 5G SA worldwide, aiming to help carriers maximize investment returns. Orange Wholesale has officially unveiled its 5G Core Network-as-a-Service (CNaaS) solution during the Mobile World Congress (MWC) 2025 in Barcelona, Spain, as the carrier looks	Uncertain	5/10	Low
Lattice Semiconductor Corporation (NASDAQ:LSCC) Shares Could Be 29% Above Their Intrinsic Value Estimate 2025-01-07 11:10:34 UTC · Earnings · Describes moves already made Model claims and source excerpt Lattice Semiconductor is estimated to be 29% overvalued with a projected fair value of US$44.34 against a current share price of US$57.27. The company is currently debt free but earnings declined over the past year. Annual earnings are forecast to grow for the next 4 years, though slower than the American market. Key Insights The projected fair value for Lattice Semiconductor is US$44.34 based on 2 Stage Free Cash Flow to Equity Lattice Semiconductor is estimated to	Relevant	5/10	Low
Newsmax Shares Surge More Than 700% in Volatile Market Debut 2025-04-01 15:54:34 UTC · Earnings · Describes moves already made Model claims and source excerpt Newsmax stock surged over 720% on its first day of public trading, closing at $82.25 after starting at $14. Trading was halted multiple times due to volatility, with the final stop occurring 14 minutes before the close. CEO Chris Ruddy retains over 81% of voting power, maintaining control over strategic direction despite the public listing. Newsmax (NMAX, Financials) saw its stock price skyrocket by more than 720% on its first day of public trading Monday, marking one of the most	Not relevant	5/10	High
Is Woodside Energy (WDS) Among the Best Australian Stocks to Buy Now? 2025-01-27 13:09:52 UTC · Earnings · Describes moves already made Model claims and source excerpt Vanguard forecasts Australia's economy will grow 2% year over year by the end of 2025, with core inflation expected to reach 2.5%. The Reserve Bank of Australia has kept its policy rate target unchanged at 4.35% since December 10, with Vanguard expecting no rate cuts until Q2 2025. Woodside Energy Group Ltd (WDS) ranks 4th on a list of best Australian stocks to buy now, based on hedge fund holdings as of Q3 2024. We recently compiled a list of the 10 Best Australian Stocks to Buy Now. In this article, we are going to take a look at	Relevant	5/10	Uncertain
The Simply Good Foods Company's (NASDAQ:SMPL) Stock On An Uptrend: Could Fundamentals Be Driving The Momentum? 2025-01-31 13:02:49 UTC · Earnings · Describes moves already made Model claims and source excerpt Simply Good Foods stock rose 12% over the past three months. The company's return on equity was 8.0% for the trailing twelve months to November 2024, below the industry average of 10%. Net income grew at a 37% rate over the last five years, outperforming the industry growth rate of 6.3%. Simply Good Foods' (NASDAQ:SMPL) stock is up by a considerable 12% over the past three months. Given that stock prices are usually aligned with a	Not relevant	5/10	Low
Celsius Holdings, Inc. (CELH): Jim Cramer’s Take on the Energy Drink Giant’s Future 2025-02-25 15:20:09 UTC · Earnings · Describes moves already made Model claims and source excerpt Celsius Holdings shares dropped 49% over the past year due to weak consumer spending and Pepsi cutting back orders, but jumped 27% in February after Q4 revenue and earnings beat estimates. Jim Cramer praised Celsius Holdings' acquisition of Alani Nutrition, calling the numbers spectacular and describing CEO Jonathan Fieldly as a great long-term performer. Jim Cramer expressed skepticism regarding the Trump administration's Department of Justice taking action against large healthcare insurers, arguing that legislation is required for meaningful change rather than lawsuits. Over the past year, the shares are down by a whopping 49% as the firm suffers from weak consumer spending. Celsius Holdings, Inc. (NASDAQ:CELH) has	Not relevant	5/10	Uncertain
Should You Buy Bank of America While It's Below $50? 2025-02-12 10:45:00 UTC · Earnings · Describes moves already made Model claims and source excerpt Bank of America shares have risen 88% since late October 2023 but remain 14% below their all-time peak. Bank of America's net interest income decreased year over year in 2024 after increasing between 2021 and 2023. The stock trades at a price-to-book ratio of 1.3, which is higher than its average over the past one, three, five, and 10 years. Bank of America (NYSE: BAC) has done a great job at winning over investors in recent times. The second-largest bank in the U.S. by assets	Not relevant	5/10	Low
NVIDIA Corporation (NVDA) Maintains ‘Buy’ Rating as AI Demand Fuels Growth Potential 2025-02-26 10:44:24 UTC · Earnings · Describes moves already made Model claims and source excerpt Evercore ISI analyst Mark Lipacis maintained a 'Buy' rating on NVIDIA (NVDA) with a $190 price target, citing undervaluation relative to historical P/E averages. Hyperscalers continue to prefer NVIDIA's platform due to robust ecosystem and GPU demand outstripping supply, mitigating concerns over Blackwell product ramp delays. The emergence of DeepSeek's R1 model triggered a $1 trillion-plus sell-off in global equities last month, while Goldman Sachs anticipates positive economic impact from AI adoption in China starting next year. since May 2014, beating its benchmark by 218 percentage points (see more details here). A close-up of a colorful high-end graphics card being plugged in	Relevant	5/10	Low
Wall Street increases range of estimated new car price hike due to tariffs 2025-04-01 11:21:02 UTC · Macro · Describes moves already made Model claims and source excerpt Wedbush Securities increased its estimate for new vehicle price hikes to offset 25% tariffs to between $5,000 and $15,000. Ferrari announced it will raise prices by up to 10% on models imported to the US after April 2. Wedbush estimates the tariffs could cost the broad auto industry $100 billion to $125 billion, eroding demand on Day 1. A Wall Street firm is increasing its estimate on how much carmakers will likely have to raise the prices of new vehicles to offset the	Relevant	5/10	Uncertain

Path to a testable risk filter

1. Corpus and timestampsImmutable versions, publication/update times, partial-text exclusions and availability audit.

2. Market labelsMassive minutes: future range and extreme moves after 09:25 / 12:00 New York. No future information in features.

3–4. Filter and scoreThe present pilot: check relevance, structured claims and systematic model errors before scaling.

5. Events and news flowCausal deduplication; new events, repeated themes and the accumulation of warnings.

6. Baseline comparisonsMarket-only, news-only and combined predictors. Probabilities come from out-of-sample evidence.

7. Time-separated validation2025 development and the frozen H6/H2 check on2026 are complete. New candidates need a separately reserved period or a forward forecast log;2026 cannot be reused as an untouched test.

8. Trading usefulnessWorst 3/5/10 sessions, missed profitable days, drawdown and false alarms at a fixed warning budget.

An important limitationModern models may remember 2025 events. Historical testing needs non-LLM controls and a frozen forward forecast log.

Reproducibility

Prompt: news_features_v3. Selection manifest SHA-256: 4b22a788b1c6f4143d5e9378c0368e52de7a8788efbb2a0dcd8c46f21f9669a3.

Only existing text is used; no publisher requests or model calls are made by this page. Reloading shows the latest saved progress.