CISD Barrier Analysis Suite

A high-performance analysis engine for testing CISD (Change in State of Delivery) patterns across NQ and ES.

This tool evaluates the "run rate" of the CISD pattern using a strict Barrier Problem approach: Does the price hit the target (CISD High/Low) before hitting the stop (opposite side) within the lookahead window?

Swing SMT confirmation is now backed by the local SMT library at /mnt/e/backup/code/Finance/Misc/SMT. The matching rule is left-only: a CISD at bar t counts when a same-direction Swing SMT was created on t, t-1, or t-2.

Planned future research ideas live in docs/research_backlog.md. The current research surface also includes standalone analyses for CISD-linked FVG creation/hold, sweep confirmation, and direction-specific swing-position buckets.

Key Findings

All figures use barrier logic: target hit before stop, lookahead = 2 bars.

⚠ Methodology correction — 2026-09-29 (quick task 260929-mkg). Read Corrected Null & Leakage Fixes before trusting any badge below. Every ✓ CONFIRMED / corrected / wf-robust / CORRECTED-BAR CLEARED badge in this README was earned against a coin-flip null (rate = 0.5). Because the barrier's target and stop are the CISD candle's own extremes, the hit rate is set mostly by where the close sits inside that candle (31% → 81% across close-location quintiles on NQ 1H), so those badges largely certify barrier geometry, not an edge. Separately, the FVG, CISD-bar-swing, FVG-size, FVG-interaction and post-CISD Reading-B buckets read bars inside their own scoring window and are retracted in their published form. The tables below are kept for provenance; the geo_* manifest columns are the corrected evidence. 2026-09-30: the corrected null now also controls the ATR regime and uses effective n (day/week clusters, pooled NQ+ES, one test per bucket). Discovery verdicts were frozen and then tested on a fresh 8-month holdout: 51 of 52 testable significant buckets kept their sign, and lifts shrank ~15–20%. See Corrected results.

How to read these tables

All rates are discovery-slice rates (oldest ~70% of data; OOS boundary = 2024-04-30). Full-history rates are not reported — every number went through the validation harness.

• N = discovery sample size
• 95% CI = Wilson score interval on the discovery slice (pure-Python, no dependencies)
• ✓ CONFIRMED = eligible (n ≥ 50 on discovery) and the OOS rate held the same side of 50%
• ✗ NOT CONFIRMED = eligible but the OOS rate did not hold — a retired hypothesis, not a finding
• below-n / not a finding = discovery n < 50; shown so the bucket can be watched as n grows

below-n and ✗ NOT CONFIRMED are different states. below-n means the bucket never had enough data to evaluate — it may be real or may not. ✗ NOT CONFIRMED means the bucket was fully evaluated and the edge did not replicate out-of-sample; it should not be published as a finding.

---

1. Baseline (Basic Run Rate)

Timeframe	Instrument	Direction	Discovery Rate	N	95% CI	Verdict
Daily	NQ	Bullish	61.0%	292	[55.3–66.4%]	✓ CONFIRMED
Daily	NQ	Bearish	59.1%	279	[53.3–64.7%]	✗ NOT CONFIRMED
Daily	ES	Bullish	57.7%	293	[52.0–63.2%]	✓ CONFIRMED
Daily	ES	Bearish	55.3%	293	[49.6–60.9%]	✗ NOT CONFIRMED
4H	NQ	Bullish	54.9%	1471	[52.4–57.5%]	✓ CONFIRMED
4H	NQ	Bearish	50.2%	1465	[47.7–52.8%]	✓ CONFIRMED
4H	ES	Bullish	57.1%	1463	[54.5–59.6%]	✓ CONFIRMED
4H	ES	Bearish	51.9%	1458	[49.3–54.4%]	✓ CONFIRMED
1H	NQ	Bullish	61.4%	5462	[60.1–62.7%]	✓ CONFIRMED
1H	NQ	Bearish	57.3%	5455	[56.0–58.6%]	✓ CONFIRMED
1H	ES	Bullish	62.7%	5235	[61.4–64.0%]	✓ CONFIRMED
1H	ES	Bearish	58.5%	5175	[57.1–59.8%]	✓ CONFIRMED
15min	NQ	Bullish	62.4%	21518	[61.7–63.0%]	✓ CONFIRMED
15min	NQ	Bearish	59.1%	21400	[58.5–59.8%]	✓ CONFIRMED
15min	ES	Bullish	62.4%	20168	[61.7–63.0%]	✓ CONFIRMED
15min	ES	Bearish	59.9%	20006	[59.2–60.6%]	✓ CONFIRMED

Key findings: Bullish bias holds OOS at all timeframes. Bearish Daily failed OOS for both instruments — Daily bearish edges should not be treated as established findings. 4H is the weakest timeframe; 1H and 15min are the most consistent.

---

2. Wick Position (Strongest Edge)

A CISD that closes past the previous wick is the single most reliable filter.

Timeframe	Instrument	Direction	Bucket	Discovery Rate	N	95% CI	Verdict
Daily	NQ	Bullish	past_wick	77.7%	94	[68.2–84.9%]	✓ CONFIRMED
Daily	NQ	Bullish	within_wick	53.0%	198	[46.1–59.9%]	✓ CONFIRMED
Daily	NQ	Bearish	past_wick	75.0%	100	[65.7–82.5%]	✓ CONFIRMED
Daily	NQ	Bearish	within_wick	50.3%	179	[43.0–57.5%]	✗ NOT CONFIRMED
Daily	ES	Bullish	past_wick	70.1%	107	[60.8–77.9%]	✓ CONFIRMED
Daily	ES	Bullish	within_wick	50.5%	186	[43.4–57.6%]	✓ CONFIRMED
Daily	ES	Bearish	past_wick	73.0%	100	[63.6–80.7%]	✓ CONFIRMED
Daily	ES	Bearish	within_wick	46.1%	193	[39.2–53.2%]	✓ CONFIRMED
4H	NQ	Bullish	past_wick	62.7%	491	[58.4–66.9%]	✓ CONFIRMED
4H	NQ	Bullish	within_wick	51.0%	980	[47.9–54.1%]	✓ CONFIRMED
4H	NQ	Bearish	past_wick	61.7%	488	[57.3–65.9%]	✓ CONFIRMED
4H	NQ	Bearish	within_wick	44.5%	977	[41.4–47.7%]	✓ CONFIRMED
4H	ES	Bullish	past_wick	67.2%	466	[62.8–71.3%]	✓ CONFIRMED
4H	ES	Bullish	within_wick	52.4%	997	[49.3–55.4%]	✓ CONFIRMED
4H	ES	Bearish	past_wick	62.3%	443	[57.7–66.7%]	✓ CONFIRMED
4H	ES	Bearish	within_wick	47.3%	1015	[44.2–50.4%]	✓ CONFIRMED
1H	NQ	Bullish	past_wick	71.9%	1669	[69.7–74.0%]	✓ CONFIRMED
1H	NQ	Bullish	within_wick	56.8%	3793	[55.2–58.3%]	✓ CONFIRMED
1H	NQ	Bearish	past_wick	68.9%	1567	[66.6–71.2%]	✓ CONFIRMED
1H	NQ	Bearish	within_wick	52.6%	3888	[51.1–54.2%]	✓ CONFIRMED
1H	ES	Bullish	past_wick	73.3%	1602	[71.1–75.5%]	✓ CONFIRMED
1H	ES	Bullish	within_wick	58.1%	3633	[56.4–59.6%]	✓ CONFIRMED
1H	ES	Bearish	past_wick	69.3%	1496	[66.9–71.5%]	✓ CONFIRMED
1H	ES	Bearish	within_wick	54.1%	3679	[52.5–55.7%]	✓ CONFIRMED
15min	NQ	Bullish	past_wick	72.5%	6281	[71.4–73.6%]	✓ CONFIRMED
15min	NQ	Bullish	within_wick	58.2%	15237	[57.4–58.9%]	✓ CONFIRMED
15min	NQ	Bearish	past_wick	70.5%	6114	[69.4–71.7%]	✓ CONFIRMED
15min	NQ	Bearish	within_wick	54.6%	15286	[53.8–55.4%]	✓ CONFIRMED
15min	ES	Bullish	past_wick	73.9%	5565	[72.7–75.0%]	✓ CONFIRMED
15min	ES	Bullish	within_wick	58.0%	14603	[57.2–58.8%]	✓ CONFIRMED
15min	ES	Bearish	past_wick	73.3%	5570	[72.2–74.5%]	✓ CONFIRMED
15min	ES	Bearish	within_wick	54.7%	14436	[53.9–55.5%]	✓ CONFIRMED

Past-wick closure is confirmed across all timeframes and both instruments. Spread between past-wick and within-wick is ~15–25pp at all timeframes. Daily NQ bearish within-wick (50.3%, n=179) did not survive OOS — within-wick bearish setups on the Daily are not an established finding.

---

3. Combined Wick × Consecutive Candles

Combining past-wick closure with 2–3 consecutive opposite candles consistently yields the highest discovery hit rates. Many Daily combined buckets are below-n / not a finding due to small sample sizes; the 1H and 15min buckets are the most fully evaluated.

Daily — selected confirmed and not-confirmed buckets (many Daily cells are below-n):

Timeframe	Instrument	Direction	Bucket	Discovery Rate	N	95% CI	Verdict
Daily	NQ	Bullish	1c_within_wick	54.1%	111	[44.8–63.0%]	✓ CONFIRMED
Daily	NQ	Bearish	1c_within_wick	51.2%	84	[40.7–61.6%]	✗ NOT CONFIRMED
Daily	NQ	Bearish	3c_within_wick	54.5%	55	[41.5–66.9%]	✗ NOT CONFIRMED
Daily	ES	Bullish	1c_past_wick	72.7%	55	[59.8–82.7%]	✓ CONFIRMED
Daily	ES	Bullish	1c_within_wick	48.9%	90	[38.8–59.0%]	✗ NOT CONFIRMED
Daily	ES	Bullish	2c_within_wick	54.4%	57	[41.6–66.6%]	✗ NOT CONFIRMED
Daily	ES	Bearish	1c_past_wick	71.4%	56	[58.5–81.6%]	✓ CONFIRMED
Daily	ES	Bearish	1c_within_wick	43.3%	90	[33.6–53.6%]	✓ CONFIRMED
Daily	ES	Bearish	3c_within_wick	50.0%	54	[37.1–62.9%]	✓ CONFIRMED

4H — all buckets:

Instrument	Direction	Bucket	Discovery Rate	N	95% CI	Verdict
NQ	Bullish	1c_past_wick	60.9%	261	[54.9–66.6%]	✓ CONFIRMED
NQ	Bullish	1c_within_wick	52.6%	532	[48.4–56.8%]	✓ CONFIRMED
NQ	Bullish	2c_past_wick	59.7%	129	[51.1–67.8%]	✓ CONFIRMED
NQ	Bullish	2c_within_wick	49.6%	238	[43.3–55.9%]	✗ NOT CONFIRMED
NQ	Bullish	3c_past_wick	71.3%	101	[61.8–79.2%]	✓ CONFIRMED
NQ	Bullish	3c_within_wick	48.6%	210	[41.9–55.3%]	✓ CONFIRMED
NQ	Bearish	1c_past_wick	66.8%	238	[60.6–72.5%]	✓ CONFIRMED
NQ	Bearish	1c_within_wick	41.8%	474	[37.4–46.3%]	✓ CONFIRMED
NQ	Bearish	2c_past_wick	55.9%	111	[46.6–64.7%]	✓ CONFIRMED
NQ	Bearish	2c_within_wick	46.4%	252	[40.4–52.6%]	✓ CONFIRMED
NQ	Bearish	3c_past_wick	57.6%	139	[49.2–65.5%]	✓ CONFIRMED
NQ	Bearish	3c_within_wick	47.8%	251	[41.7–53.9%]	✓ CONFIRMED
ES	Bullish	1c_past_wick	64.0%	272	[58.1–69.4%]	✓ CONFIRMED
ES	Bullish	1c_within_wick	54.7%	525	[50.4–58.9%]	✓ CONFIRMED
ES	Bullish	2c_past_wick	67.9%	109	[58.6–75.9%]	✓ CONFIRMED
ES	Bullish	2c_within_wick	50.4%	248	[44.2–56.6%]	✓ CONFIRMED
ES	Bullish	3c_past_wick	76.5%	85	[66.4–84.2%]	✓ CONFIRMED
ES	Bullish	3c_within_wick	49.1%	224	[42.6–55.6%]	✗ NOT CONFIRMED
ES	Bearish	1c_past_wick	59.1%	230	[52.7–65.3%]	✓ CONFIRMED
ES	Bearish	1c_within_wick	45.9%	525	[41.7–50.2%]	✓ CONFIRMED
ES	Bearish	2c_past_wick	65.5%	87	[55.1–74.7%]	✓ CONFIRMED
ES	Bearish	2c_within_wick	49.2%	238	[42.9–55.5%]	✓ CONFIRMED
ES	Bearish	3c_past_wick	65.9%	126	[57.2–73.6%]	✓ CONFIRMED
ES	Bearish	3c_within_wick	48.4%	252	[42.3–54.6%]	✓ CONFIRMED

1H and 15min combined buckets: All 48 buckets (24 each) are ✓ CONFIRMED. Past-wick + 2c or 3c at 1H/15min consistently reaches 70–75% discovery rates and holds OOS.

Note on Daily combined buckets: Most Daily combined cells fall below-n / not a finding (discovery n < 50). Of the 24 Daily combined buckets with n ≥ 50, 8 are ✗ NOT CONFIRMED (concentrated in within-wick bearish and ES bullish within-wick). Daily past-wick + any consecutive count confirms for both instruments where n ≥ 50.

---

4. Stricter CISD (Significance Test)

Requiring the close to exceed the previous candle's High/Low (not just the close) gives a consistent +5–8pp lift over baseline. All 16 buckets are ✓ CONFIRMED.

Timeframe	Instrument	Direction	Discovery Rate	N	95% CI	Verdict
Daily	NQ	Bullish	68.4%	329	[63.2–73.2%]	✓ CONFIRMED
Daily	NQ	Bearish	65.3%	259	[59.3–70.8%]	✓ CONFIRMED
Daily	ES	Bullish	68.2%	333	[62.9–72.9%]	✓ CONFIRMED
Daily	ES	Bearish	65.5%	261	[59.6–71.0%]	✓ CONFIRMED
4H	NQ	Bullish	60.3%	1619	[57.9–62.6%]	✓ CONFIRMED
4H	NQ	Bearish	58.3%	1332	[55.7–60.9%]	✓ CONFIRMED
4H	ES	Bullish	63.1%	1567	[60.7–65.5%]	✓ CONFIRMED
4H	ES	Bearish	59.7%	1269	[56.9–62.3%]	✓ CONFIRMED
1H	NQ	Bullish	65.8%	5644	[64.6–67.0%]	✓ CONFIRMED
1H	NQ	Bearish	63.6%	4686	[62.2–64.9%]	✓ CONFIRMED
1H	ES	Bullish	66.6%	5546	[65.4–67.9%]	✓ CONFIRMED
1H	ES	Bearish	64.8%	4594	[63.4–66.1%]	✓ CONFIRMED
15min	NQ	Bullish	67.5%	21256	[66.9–68.2%]	✓ CONFIRMED
15min	NQ	Bearish	66.2%	19161	[65.5–66.9%]	✓ CONFIRMED
15min	ES	Bullish	68.6%	20307	[67.9–69.2%]	✓ CONFIRMED
15min	ES	Bearish	67.7%	18387	[66.9–68.3%]	✓ CONFIRMED

---

5. Consecutive Opposite Candles (Markov)

Consecutive candle count alone has weak and inconsistent predictive value at Daily and partially at 4H. At 1H and 15min, all consecutive buckets are confirmed.

Daily — all buckets (many not-confirmed):

Instrument	Direction	Bucket	Discovery Rate	N	95% CI	Verdict
NQ	Bullish	1_consecutive	61.0%	159	[53.3–68.2%]	✓ CONFIRMED
NQ	Bullish	2_consecutive	63.4%	71	[51.8–73.6%]	✗ NOT CONFIRMED
NQ	Bullish	3_consecutive	58.1%	62	[45.7–69.5%]	✓ CONFIRMED
NQ	Bearish	1_consecutive	58.3%	127	[49.6–66.5%]	✗ NOT CONFIRMED
NQ	Bearish	2_consecutive	57.6%	66	[45.6–68.8%]	✗ NOT CONFIRMED
NQ	Bearish	3_consecutive	61.6%	86	[51.1–71.2%]	✗ NOT CONFIRMED
ES	Bullish	1_consecutive	57.9%	145	[49.8–65.7%]	✓ CONFIRMED
ES	Bullish	2_consecutive	62.4%	85	[51.7–71.9%]	✗ NOT CONFIRMED
ES	Bullish	3_consecutive	50.8%	63	[38.8–62.7%]	✓ CONFIRMED
ES	Bearish	1_consecutive	54.1%	146	[46.0–61.9%]	✗ NOT CONFIRMED
ES	Bearish	2_consecutive	59.2%	71	[47.5–69.8%]	✗ NOT CONFIRMED
ES	Bearish	3_consecutive	53.9%	76	[42.8–64.6%]	✗ NOT CONFIRMED

4H — mixed results; 6 of 12 not-confirmed (bearish ES and some bullish NQ).

1H and 15min — all 24 buckets confirmed. Hit rates are flat across 1–3 consecutive candles (within ~3pp of baseline). The edge from Markov segmentation alone is small; it only emerges clearly when combined with wick position (see §3).

---

6. Candle Body Size vs ATR & Cross-Tab

(See standalone charts: CandleSize_All_Timeframes.png, SizeCross_All_Timeframes.png)

CISD candles with a body ≥ 0.5× ATR(14) show meaningfully higher hit rates than smaller candles. The cross-tab reveals a clear quadrant structure: Big CISD + Small prev = strongest; Small CISD + Big prev = weakest. The pattern is robust at 1H and 15min (all buckets confirmed); some Daily and 4H cells are below-n / not a finding or ✗ NOT CONFIRMED.

Representative confirmed candle_size buckets (1H, NQ Bullish):

Bucket	Discovery Rate	N	95% CI	Verdict
< 0.5× ATR	55.6%	3448	[53.9–57.2%]	✓ CONFIRMED
0.5–1× ATR	68.2%	1225	[65.6–70.8%]	✓ CONFIRMED
1–1.5× ATR	72.6%	503	[68.5–76.3%]	✓ CONFIRMED
> 1.5× ATR	82.7%	284	[77.9–86.7%]	✓ CONFIRMED

Size cross-tab confirmed quadrant (1H, NQ Bullish):

Bucket	Discovery Rate	N	95% CI	Verdict
Big CISD / Small prev	76.7%	593	[73.2–80.0%]	✓ CONFIRMED
Big CISD / Big prev	74.7%	194	[68.2–80.3%]	✓ CONFIRMED
Small CISD / Small prev	58.7%	4098	[57.1–60.2%]	✓ CONFIRMED
Small CISD / Big prev	60.7%	575	[56.6–64.6%]	✓ CONFIRMED

Daily and 4H notes: Many "Big CISD" Daily buckets fall below-n / not a finding (n < 50). At 4H NQ bullish, the < 0.5× ATR bucket (n=950) is ✗ NOT CONFIRMED. All 15min candle_size and size_cross buckets are ✓ CONFIRMED, confirming the ATR-segmentation pattern across large samples.

---

7. Volume Ratio

(See Volume_All_Timeframes.png)

Volume ratio (CISD candle vs previous candle) has negligible impact on outcomes at 1H and 15min — hit rates are stable across all volume buckets and all 32 1H/15min buckets are ✓ CONFIRMED. At Daily and 4H, some buckets are ✗ NOT CONFIRMED (particularly 1x–1.5x bearish Daily and several 4H cells), and a handful are below-n / not a finding due to small n in the 1.5x–2.5x bin.

Representative confirmed volume buckets (1H, NQ Bullish):

Bucket	Discovery Rate	N	95% CI	Verdict
< 1× (lower vol)	60.5%	3058	[58.8–62.2%]	✓ CONFIRMED
1×–1.5×	62.7%	1316	[60.0–65.3%]	✓ CONFIRMED
1.5×–2.5×	61.7%	741	[58.1–65.1%]	✓ CONFIRMED
> 2.5× (spike)	63.4%	347	[58.2–68.3%]	✓ CONFIRMED

The near-flat distribution across all four buckets at 1H/15min confirms that volume ratio does not add meaningful edge as a standalone filter on CISD barrier outcomes at those timeframes.

---

8. Swing SMT Confirmation

(See SMT_CISD_All_Timeframes.png)

A co-occurring same-direction Swing SMT (divergence between NQ and ES swing highs/lows, lookback=20) is used as a confirmation filter. As of v2.0 the w/ SMT population is corrected — see "SMT Invalidation Honesty (v2.0)" below for what changed and why. Results vary sharply by timeframe.

Timeframe	Instrument	Direction	w/ SMT Rate	w/ SMT N	95% CI	no SMT Rate	Δ	Verdict (w/ SMT)
Daily	NQ	Bullish	69.2%	39	[53.6–81.4%]	60.5%	+8.7pp	below-n / not a finding
Daily	NQ	Bearish	64.3%	28	[45.8–79.3%]	57.5%	+6.8pp	below-n / not a finding
Daily	ES	Bullish	57.1%	42	[42.2–70.9%]	59.1%	−1.9pp	below-n / not a finding
Daily	ES	Bearish	51.4%	35	[35.6–67.0%]	55.4%	−4.0pp	below-n / not a finding
4H	NQ	Bullish	59.0%	156	[51.1–66.4%]	54.9%	+4.1pp	✓ CONFIRMED
4H	NQ	Bearish	50.3%	171	[42.9–57.7%]	50.1%	+0.2pp	✓ CONFIRMED
4H	ES	Bullish	64.5%	152	[56.6–71.6%]	56.7%	+7.8pp	✓ CONFIRMED
4H	ES	Bearish	52.2%	161	[44.5–59.7%]	51.1%	+1.0pp	✓ CONFIRMED
1H	NQ	Bullish	59.0%	503	[54.7–63.3%]	61.2%	−2.1pp	✓ CONFIRMED
1H	NQ	Bearish	54.9%	565	[50.7–58.9%]	57.5%	−2.7pp	✓ CONFIRMED
1H	ES	Bullish	60.4%	472	[55.9–64.7%]	62.6%	−2.3pp	✓ CONFIRMED
1H	ES	Bearish	55.3%	535	[51.1–59.5%]	58.8%	−3.4pp	✓ CONFIRMED
15min	NQ	Bullish	64.9%	1887	[62.7–67.0%]	61.9%	+2.9pp	✓ CONFIRMED
15min	NQ	Bearish	61.2%	1965	[59.0–63.4%]	58.6%	+2.6pp	✓ CONFIRMED
15min	ES	Bullish	63.9%	1810	[61.7–66.1%]	62.1%	+1.9pp	✓ CONFIRMED
15min	ES	Bearish	60.5%	1884	[58.3–62.7%]	59.7%	+0.8pp	✓ CONFIRMED

The invalidation fix also surfaces the expired SMT bucket — CISDs that matched a same-direction Swing SMT in the [t-2, t] window, but that SMT was already invalidated (broken_ts <= t) by the time the CISD formed. These were previously mis-credited as w/ SMT; they are now reported separately and are not part of the corrected w/ SMT rate above:

Timeframe	Instrument	Direction	expired SMT Rate	expired SMT N
Daily	NQ	Bullish	40.0%	10
Daily	NQ	Bearish	72.2%	18
Daily	ES	Bullish	22.2%	9
Daily	ES	Bearish	61.1%	18
4H	NQ	Bullish	46.0%	63
4H	NQ	Bearish	53.3%	60
4H	ES	Bullish	47.6%	63
4H	ES	Bearish	65.6%	61
1H	NQ	Bullish	71.6%	225
1H	NQ	Bearish	58.5%	260
1H	ES	Bullish	69.9%	219
1H	ES	Bearish	59.9%	252
15min	NQ	Bullish	66.3%	816
15min	NQ	Bearish	65.0%	857
15min	ES	Bullish	65.2%	747
15min	ES	Bearish	62.2%	788

Key takeaways (corrected, v2.0):

• Daily SMT: all four w/ SMT buckets are now below-n / not a finding (n < 50 in every combo) — the two ES buckets that were previously eligible (n=51, n=53) dropped below the n=50 gate once the expired SMT split removed already-dead matches from the w/ SMT population. Daily SMT carries no finding at either the before-fix or after-fix bar.
• 4H SMT: all four w/ SMT buckets are confirmed post-fix. ES Bearish flips from ✗ NOT CONFIRMED to ✓ CONFIRMED — the pre-fix bucket's discovery rate (55.9%) held on OOS at 46.7% (opposite side), but the corrected discovery rate (52.2%) now holds the same side as the corrected OOS rate (51.5%). This flip is a direct, documented consequence of the invalidation fix, not a new discovery.
• 1H SMT: all four w/ SMT buckets remain confirmed; rates shift down modestly (−2 to −4pp vs the pre-fix numbers) now that expired matches are excluded, but the same-side-of-50% OOS confirmation is unaffected.
• 15min remains the strongest timeframe: all four buckets confirmed, +0.8–2.9pp over no-SMT with large discovery samples (n=1,810–1,965).
• Scope note: the w/ SMT & survived / w/ SMT & broke split (new in v2.0) is a diagnostic only — the aggregate w/ SMT population above is never filtered on survival, since that would be hindsight (an outcome only knowable after the barrier window closes). See output/smt_invalidation_report.csv and SMT_Role_All_Timeframes.png / SMT_BlockSize_All_Timeframes.png / SMT_InBlock_All_Timeframes.png for the new smt_role / smt_block_size / smt_in_block validated buckets (n + Wilson CI + BH-FDR + walk-forward), which are separate analyses, not sub-splits of smt_cisd.

---

SMT Invalidation Honesty (v2.0)

Phase 9 fixes a bug in the Swing SMT confirmation tagging: a matched same-direction SMT was credited as w/ SMT even when it had already been invalidated by the CISD bar t (broken_ts <= t). Because sub-bar ordering between an intrabar invalidation and the CISD close is unknowable, an SMT whose broken_ts lands exactly on bar t is now treated conservatively as dead — the CISD is reclassified into a new expired SMT bucket instead of w/ SMT. The swing_smt_tag split is therefore three-way as of v2.0: w/ SMT (still valid at t) / expired SMT (matched but dead by t) / no SMT (no same-direction match at all).

This changes the previously-published w/ SMT rate and n for every timeframe/instrument/direction combo — always a reduction in n (dead matches leave the bucket) and a rate shift in either direction depending on how the removed expired matches performed. Per D-09a, the "before" numbers below are the currently-published (pre-fix) §8 values already on record — no separate recompute or snapshot ceremony was needed. "After" is the corrected, regenerated discovery manifest. The table is generated by scripts/build_smt_invalidation_report.py from output/smt_invalidation_report.csv:

Timeframe	Instrument	Direction	Before Rate	Before N	After Rate	After N	Δ Rate	Δ N
Daily	NQ	Bullish	63.3%	49	69.2%	39	+6.0pp	−10
Daily	NQ	Bearish	67.4%	46	64.3%	28	−3.1pp	−18
Daily	ES	Bullish	51.0%	51	57.1%	42	+6.2pp	−9
Daily	ES	Bearish	54.7%	53	51.4%	35	−3.3pp	−18
4H	NQ	Bullish	55.3%	219	59.0%	156	+3.7pp	−63
4H	NQ	Bearish	51.1%	231	50.3%	171	−0.8pp	−60
4H	ES	Bullish	59.5%	215	64.5%	152	+4.9pp	−63
4H	ES	Bearish	55.9%	222	52.2%	161	−3.7pp	−61
1H	NQ	Bullish	62.9%	728	59.0%	503	−3.9pp	−225
1H	NQ	Bearish	56.0%	825	54.9%	565	−1.1pp	−260
1H	ES	Bullish	63.4%	691	60.4%	472	−3.0pp	−219
1H	ES	Bearish	56.8%	787	55.3%	535	−1.5pp	−252
15min	NQ	Bullish	65.3%	2703	64.9%	1887	−0.4pp	−816
15min	NQ	Bearish	62.4%	2822	61.2%	1965	−1.1pp	−857
15min	ES	Bullish	64.3%	2557	63.9%	1810	−0.4pp	−747
15min	ES	Bearish	61.0%	2672	60.5%	1884	−0.5pp	−788

What changed, in plain terms:

• Every combo loses n (10–857 CISDs move out of w/ SMT into expired SMT) — the previously mis-credited already-dead-SMT population is now visible as its own bucket instead of silently inflating w/ SMT.
• Rate moves are mixed (−3.9pp to +6.2pp) depending on how the removed expired-SMT matches happened to run relative to the still-valid ones — there is no systematic direction to the correction, consistent with fixing a labeling bug rather than discovering a new effect.
• Two Daily ES buckets drop from eligible (n≥50) to below-n purely because of the n reduction — not because their behavior changed.
• One verdict flip: 4H ES Bearish moves from ✗ NOT CONFIRMED (pre-fix) to ✓ CONFIRMED (corrected) — see the §8 key takeaways above for the discovery/OOS numbers behind the flip.
• Every non-smt_* analysis is confirmed byte-value unchanged by the regeneration's own drift check (build_non_smt_drift in scripts/build_smt_invalidation_report.py, run against the regenerated output/validation_manifest_discovery.csv): it compares every non-SMT (analysis, timeframe, instrument, direction, bucket) row's rate/n before vs after and exits non-zero if anything moved. The regen that produced this table's "after" numbers reported zero drifted rows — the other 15 published analyses (Baseline, Wick Position, Combined, Stricter CISD, Consecutive Candles, Candle Size, Volume, CISD FVG, FVG Hold, CISD FVG Interaction, Sweep, SSSF Swing, and both post-CISD studies) moved by exactly nothing (D-09a).

---

Visual Reports

Per-Timeframe (5 core analyses):

Standalone — All Timeframes side-by-side:

---

Usage

Ensure data is in .parquet format in the data/ directory (nq_1m.parquet, es_1m.parquet).

Forward Returns Explorer

Rebuild the interactive forward-returns report with:

python3 scripts/build_forward_returns.py

The page groups research into three families:

• core: SMT, size-vs-prev-body, wick position, and consecutive-candle buckets.
• fvg: CISD-linked FVG creation and hold-state buckets.
• structure: sweep confirmation plus the previous-bar and CISD-bar swing-position buckets.

Run all analyses (4 per-TF PNGs + CSVs + 9 standalone PNGs):

python cisd_analysis.py

Run specific models:

python cisd_analysis.py cisd_fvg fvg_hold cisd_fvg_interaction sweep sssf_swing

The current CLI surface is limited to the analysis keys listed below.

Evaluation Models (All Barrier-Based)

Key	Model Name	Description
basic	Basic Run Rate	Baseline success rate for all CISD events.
mc	Markov Segmentation	Buckets by consecutive opposite-direction candles preceding the CISD.
significance	Stricter CISD	CISD must close past the previous candle's High (Bullish) or Low (Bearish).
wick	Wick Position	Split by whether the close was past the wick or within it.
combined	Wick x Markov	Cross-tab of wick position and consecutive candle count.
volume	Volume Ratio	Segments by CISD volume relative to previous candle. Standalone all-TF output.
candle_size	Candle Body vs ATR	Segments by CISD body size as a multiple of ATR(14). Standalone all-TF output.
size_cross	CISD Body × Prev Body	Cross-tab: both candles vs ATR(14). Standalone all-TF output.
smt_cisd	Swing SMT Confirmation	Barrier rate split by whether a same-direction, still-valid-at-t Swing SMT co-occurs (w/ SMT / expired SMT / no SMT), plus w/ SMT & survived/w/ SMT & broke diagnostic sub-buckets. Standalone.
smt_role	Swing SMT Role	Swept vs failed-to-sweep split over the valid w/ SMT population only. Standalone.
smt_block_size	SMT Block Size vs ATR	Segments the matched-SMT population by smt_block_size_atr (reference-bar range / ATR(14)) using the candle_size bucket convention. Standalone.
smt_in_block	CISD In SMT Block	Splits the matched-SMT population by whether the CISD body is fully contained within the SMT's reference-bar range (cisd_in_block/cisd_out_block). Standalone.
cisd_fvg	CISD FVG Creation	Barrier rate split by whether the CISD is the middle candle of a same-direction FVG (mid0), the next bar is the middle candle (mid1), or no linked FVG exists. Standalone.
fvg_hold	FVG Hold	Hold rate of same-direction CISD-linked FVGs over a 10-bar window from the FVG middle candle, reported for both hold-failure modes and both mid0/mid1 buckets. Standalone.
cisd_fvg_interaction	CISD FVG Interaction	Barrier rate of the parent CISD split by whether its linked same-direction FVG later held or failed, for both failure modes and both mid0/mid1 buckets. Standalone.
sweep	Sweep Confirmation	Barrier rate split by whether a same-direction sweep of a prior swing occurred in the [t-4, t] window around the CISD. Standalone.
sssf_swing	SSSF Swing	Barrier rate split by whether candle[-1] or candle[0] is the direction-matched 3-candle swing point, with a baseline neither bucket. Standalone.

Research Tag Rules

• FVG formation uses the standard 3-candle wick-to-wick rule.
  • Bullish: left high < right low.
  • Bearish: left low > right high.
  • mid0 means the CISD bar is the middle candle; mid1 means the next bar is the middle candle.
• FVG direction must match the parent CISD direction.
• FVG hold is measured over FVG_HOLD_LOOKAHEAD = 10 bars from the FVG middle candle with two failure modes:
  • close_through_near_edge
  • wick_break_far_extreme
• Sweep confirmation is direction-specific and binary:
  • bullish CISD checks for a low sweep
  • bearish CISD checks for a high sweep
  • the sweep can occur on any bar in [t-4, t]
  • prior swing points are sourced from a 20-bar swing lookback
• Swing-position buckets use a 3-candle swing rule (N=1):
  • bullish CISDs use swing lows
  • bearish CISDs use swing highs
  • prev_bar_is_swing corresponds to candle[-1]
  • cisd_bar_is_swing corresponds to candle[0]
• When each tag is known (look-ahead discipline). mid0 FVG and cisd_bar_is_swing are only known at the close of bar t+1; mid1 FVG, "no FVG", FVG size and candle2_past_candle1_wick at the close of t+2; FVG hold state after t+1+FVG_HOLD_LOOKAHEAD. Analyses conditioning on them are scored confirm-then-enter (entry at the confirming close, window after it, setups already resolved before entry excluded) or flagged diagnostic — see Corrected Null & Leakage Fixes.
• Session tags are taken at the bar midpoint (bars are labeled by their start), so the 1H 09:00 bar — which contains the 09:30 open — is rth_open.

Validation Methodology — Harder Evidence Bar (v2.0)

Phase 6 adds two additive validation capabilities to scripts/build_validation.py. Neither changes an existing manifest column or a previously published rate — both are new columns / new sibling artifacts layered on top of the v1.0 harness (sacred chronological holdout, Wilson CI, min_n_pass gate).

• D-04 scope disclaimer: This harder corrected bar currently applies only to the two post-CISD studies, post_cisd_context and candle1_followthrough. The other 8 published Key Findings sections — Baseline, Wick Position, Combined Wick×Consecutive, Stricter CISD, Consecutive Opposite Candles, Candle Body Size, Volume Ratio, and Swing SMT — have not been re-evaluated under it.

Multiple-comparisons (FDR) correction:

• Every bucket in the discovery manifest now carries a p_value testing H0: rate = 0.5 (the fixed coin-flip null, not a bucket's own baseline rate) — a two-sided normal-approximation z-test (p_value_vs_half).
• A single global Benjamini-Hochberg step-up correction (bh_correct) is applied across every bucket, across every analysis × timeframe × instrument × direction, in one pass — never grouped per-analysis-key — because the point is to account for the full extent of data-snooping across the whole engine, not one chart at a time.
• The correction fires at the discovery-manifest stage only (validation_manifest_discovery.csv), before the sacred OOS look; it is a pre-registration-style gate on which buckets earn that look. The OOS manifest is unaffected.
• Adds bh_rank, bh_q_value, bh_significant, and corrected_pass columns to validation_manifest_discovery.csv. corrected_pass = min_n_pass AND bh_significant — the harder evidence bar requires clearing both the sample-size gate and FDR significance.

Walk-forward (rolling-window) validation:

• Run with python3 scripts/build_validation.py --walk-forward. Output: a new sibling artifact, output/validation_manifest_walkforward.csv — the existing validation_manifest_discovery.csv / validation_manifest_oos.csv files and the sacred OOS banner are never touched by this path.
• Windows are expanding / anchored: each of the 4 folds trains on all discovery history from the start through its fold boundary, then tests on the following chunk — chosen because the discovery history is finite and every available bar should be used rather than discarded.
• Fold boundaries are 4 frozen calendar dates (WALK_FORWARD_FOLDS in cisd_data.py), derived once as the 20th/40th/60th/80th-percentile dates of the discovery-region calendar (mirroring how OOS_START itself was derived and frozen) — not recomputed at runtime, and every boundary is strictly before OOS_START, so walk-forward never consumes the one sacred OOS evaluation.
• Each fold is scored by evaluate_fold(): both the anchored train window and the test chunk must independently clear MIN_N (a per-fold sample-size gate), and the test rate must confirm the train rate's same non-boundary side of 50% (an exact 50% train rate makes no directional prediction and cannot pass).
• The per-bucket aggregate wf_verdict (walk_forward_verdict()) requires a majority — strictly more than 50% — of folds to pass (wf-robust); exactly 50% is wf-fragile, not robust. below-n folds still count in the denominator.
• validation_manifest_walkforward.csv columns: analysis, timeframe, instrument, direction, bucket, fold_index, train_end, test_end, train_rate, train_n, test_rate, test_n, fold_verdict, wf_verdict.

Post-CISD Context — Corrected Re-Validation (v2.0)

This section extends the badge vocabulary with a fourth, strictly harder tier: ✓ CORRECTED-BAR CLEARED. Under the D-02 three-condition AND, a bucket must have discovery corrected_pass, walk-forward wf_verdict == 'wf-robust', and a sacred-OOS rate on the same side of 0.50 as discovery. This is stricter than ✓ CONFIRMED, which checks the OOS same-side condition without the FDR-correction and walk-forward legs. not-cleared means the tag did not clear this corrected bar; it does not replace the existing ✗ NOT CONFIRMED or below-n / not a finding states.

analysis	tag	n_buckets	n_cleared	verdict
candle1_followthrough	against_forward	16	12	cleared
candle1_followthrough	against_inwindow	16	12	cleared
candle1_followthrough	with_past_wick_forward	16	12	cleared
candle1_followthrough	with_past_wick_inwindow	16	12	cleared
candle1_followthrough	with_within_wick_forward	16	8	not-cleared
candle1_followthrough	with_within_wick_inwindow	16	9	cleared
post_cisd_context	candle2_past_candle1_wick	16	12	cleared
post_cisd_context	failed_gap_against	16	10	cleared
post_cisd_context	failed_gap_against_neither	16	10	cleared
post_cisd_context	failed_gap_against_reversal	16	10	cleared
post_cisd_context	failed_gap_flat	16	10	cleared
post_cisd_context	failed_gap_with	16	11	cleared

⚠ 2026-09-29: the *_inwindow tags are outcome-leaking (candle[1] is inside their scoring window) and are now diagnostic; candle2_past_candle1_wick read close[t+2] inside its forward window and is now scored from close[t+2]. All "cleared" verdicts here used the 0.5 null — see Corrected Null & Leakage Fixes.

For the failed_gap_against reversal-barrier reading, the mutually exclusive outcomes partition the same population. Discovery is 32.9% continuation (failed_gap_against, n=23,247), 59.2% reversal (n=23,247), and 7.9% neither (n=23,247). Sacred OOS is 34.3% continuation (n=10,706), 57.8% reversal (n=10,706), and 7.9% neither (n=10,706). The majority therefore reverse after a failed-gap-against CISD; that reversal rate is distinct from, and is not conflated with, the continuation rate in the CISD direction.

Conditioning Features — Magnitude, Session & Volume Anomaly (v2.0)

Phase 10 (RES-07) adds three economically-motivated conditioning-feature families on top of the existing binary flags, each pushed through the full corrected harness (n ≥ 50 gate, Wilson CI, one global Benjamini-Hochberg FDR correction, and walk-forward robustness) with zero harness change — every new analysis is an additive standalone compute_* / chart_* / ANALYSES entry (the Phase-9 "new feature = additive standalone analysis" pattern). All 408 new discovery buckets are published, including those that do not clear the bar — the point is to hand Phase 11's model economically-motivated, harness-validated inputs rather than let it mine noise.

Families:

1. Magnitude — continuous ATR-normalized versions of existing binary flags: wick_distance (signed distance past the prior wick, with a bin edge exactly at 0 so the old past/within split stays recoverable), sweep_depth (how far the CISD extreme pierced the swept level), fvg_size (gap width).
2. Session / time-of-day — the engine's first temporal dimension: a frozen 3-bucket rth_open (09:30–10:30 ET) / rth (10:30–16:00) / overnight (16:00–09:30) split from the tz-naive ET index, with no timezone or DST math. Registered on 15min and 1H only — a session tag on a Daily/4H bar spans multiple sessions and is economically meaningless, and would only add below-n cells that dilute the global FDR family.
3. Volume anomaly — measures beyond the already-negligible 1-bar volume ratio (which stays untouched): effort_result (within-bar volume ÷ range, baseline-free), and slot-normalized rvol / volume_zscore (same-time-of-day trailing baseline, frozen K=20, so the intraday volume profile is removed rather than measured).

Badges in this section. corrected is the discovery BH-FDR verdict (corrected_pass = min_n_pass AND bh_significant); below-n = discovery n < 50; ✗ NOT CONFIRMED = eligible but did not clear the corrected bar. OOS is the sacred-holdout rate; walk-fwd is the expanding-window wf_verdict. The strongest evidence is a bucket that is corrected-confirmed and walk-forward wf-robust and whose OOS rate holds the discovery side of 0.50. All frozen bins were shape-informed once from the discovery-slice histogram (index < OOS_START, ~225k obs) outcome-blind — chosen from the distribution, never tuned to hit rates — then frozen in-code.

Summary — 408 new discovery buckets

feature	✓ confirmed	✗ not-conf	below-n	wf-robust	wf-fragile
wick_distance	44	12	8	32	32
sweep_depth	30	7	27	16	48
fvg_size	44	0	20	16	48
effort_result	30	8	26	28	36
rvol	45	13	6	37	27
volume_zscore	47	13	4	36	28
session	24	0	0	20	4

(wf-robust / wf-fragile count each bucket once; the confirmed/not-confirmed/below-n split is the discovery BH-FDR verdict.)

Representative results

Magnitude · signed wick distance — monotonic, the standout (15min NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<-1x ATR (deep within wick)	55.5%	1880	[53.2%–57.7%]	✓ CONFIRMED	53.9%	wf-robust
-1x-0 ATR (within wick)	58.5%	13354	[57.7%–59.4%]	✓ CONFIRMED	58.0%	wf-robust
0-1x ATR (past wick)	72.2%	5790	[71.1%–73.4%]	✓ CONFIRMED	71.5%	wf-robust
>1x ATR (far past wick)	76.2%	491	[72.2%–79.7%]	✓ CONFIRMED	76.4%	wf-robust

Magnitude · FVG size (15min NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<0.5x ATR	92.5%	3913	[91.6%–93.3%]	✓ CONFIRMED	92.4%	wf-robust
0.5x-1x ATR	92.3%	888	[90.4%–93.9%]	✓ CONFIRMED	93.8%	wf-robust
1x-1.5x ATR	89.8%	283	[85.7%–92.8%]	✓ CONFIRMED	92.7%	wf-robust
>1.5x ATR	87.4%	182	[81.8%–91.4%]	✓ CONFIRMED	76.0%	wf-fragile

Magnitude · sweep depth — confirmed but flat (little gradient) (15min NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<0.5x ATR	70.5%	664	[66.9%–73.8%]	✓ CONFIRMED	71.3%	wf-robust
0.5x-1x ATR	72.1%	458	[67.8%–76.0%]	✓ CONFIRMED	70.2%	wf-robust
1x-1.5x ATR	71.5%	354	[66.6%–75.9%]	✓ CONFIRMED	68.1%	wf-robust
>1.5x ATR	66.8%	596	[62.9%–70.4%]	✓ CONFIRMED	69.1%	wf-robust

Volume · slot-relative RVOL — monotonic (15min NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<0.7x slot	58.3%	6723	[57.1%–59.5%]	✓ CONFIRMED	58.7%	wf-robust
0.7x-1x slot	63.2%	6670	[62.0%–64.3%]	✓ CONFIRMED	61.9%	wf-robust
1x-1.5x slot (elevated)	64.7%	5324	[63.4%–65.9%]	✓ CONFIRMED	63.9%	wf-robust
>1.5x slot (spike)	65.4%	2359	[63.5%–67.3%]	✓ CONFIRMED	66.2%	wf-robust

Volume · slot z-score — monotonic (15min NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<-0.5 sigma	58.8%	8093	[57.7%–59.9%]	✓ CONFIRMED	59.3%	wf-robust
-0.5-0.5 sigma	63.7%	8510	[62.7%–64.7%]	✓ CONFIRMED	62.7%	wf-robust
0.5-1.5 sigma	66.0%	2794	[64.2%–67.7%]	✓ CONFIRMED	64.7%	wf-robust
>1.5 sigma (spike)	65.3%	1679	[63.0%–67.6%]	✓ CONFIRMED	64.8%	wf-robust

Volume · effort-vs-result — noisy / TF-scale-fragmented (15min ES bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
<150	62.5%	8	[30.6%–86.3%]	below-n	68.9%	wf-fragile
150-550	65.8%	4124	[64.3%–67.2%]	✓ CONFIRMED	63.7%	wf-robust
550-1500	61.2%	8294	[60.2%–62.3%]	✓ CONFIRMED	59.5%	wf-robust
>1500	61.7%	7742	[60.6%–62.8%]	✓ CONFIRMED	61.5%	wf-robust

Session · time-of-day — rth_open > rth > overnight (1H NQ bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
rth_open	67.4%	227	[61.1%–73.2%]	✓ CONFIRMED	63.4%	wf-fragile
rth	66.1%	1143	[63.3%–68.7%]	✓ CONFIRMED	64.7%	wf-robust
overnight	59.8%	4092	[58.2%–61.2%]	✓ CONFIRMED	62.7%	wf-robust

Session · time-of-day (15min ES bullish)

bucket	disc rate	n	Wilson CI	corrected	OOS	walk-fwd
rth_open	66.3%	879	[63.1%–69.4%]	✓ CONFIRMED	69.2%	wf-robust
rth	64.8%	4838	[63.4%–66.1%]	✓ CONFIRMED	63.2%	wf-robust
overnight	61.3%	14451	[60.5%–62.1%]	✓ CONFIRMED	60.9%	wf-robust

Interpretation

⚠ Superseded 2026-09-29 — these readings were made against the 0.5 null and, for fvg_size, a leaking frame; session also used mislabeled 1H buckets. See Corrected Null & Leakage Fixes for the corrected evidence. Kept for provenance.

• wick_distance — the standout. The signed magnitude is cleanly monotonic and genuinely subsumes the binary wick split: continuation climbs deep-within → within → just-past → far-past-wick (≈55% → 58% → 72% → 76%+), holding out-of-sample and walk-forward-robust across the high-n intraday timeframes. It enriches compute_wick, it does not merely restate it.
• fvg_size. Small FVGs (< 0.5× ATR) continue ≈90–95%; the rate softens and turns walk-forward-fragile for the largest gaps (> 1.5× ATR — e.g. 15min NQ bull OOS 76%, wf-fragile). Tight gaps are the higher-continuation regime.
• session. A clean, fully-confirmed temporal ordering — rth_open > rth > overnight — on both 1H and 15min: the opening-drive hour carries the highest CISD continuation. The engine's first temporal edge.
• rvol / volume_zscore. A modest but consistent, monotonic, walk-forward-robust anomaly signal — higher slot-relative volume → higher continuation — and the two independent measures agree.
• sweep_depth. Confirmed but flat across depth (≈67–72%, no gradient): the pierce-depth magnitude adds little beyond the binary sweep flag it enriches.
• effort_result. The weakest and noisiest — its raw baseline-free ratio has a scale that differs sharply across timeframes (D-07), fragmenting the fixed bins so Daily/4H buckets are mostly below-n while the intraday buckets confirm without a strong gradient.

Honesty caveats (D-11)

• OHLCV-only limitation. These volume features are unsigned effort / anomaly proxies computed from OHLCV alone. No signed or delta order flow is available on this data, so none of them can distinguish buying from selling pressure — they measure how much volume/effort, never which side. Read them as activity-intensity conditioning, not order flow.
• Slot-normalization is partial. The slot baseline (same time-of-day, trailing K=20) neutralizes most but not necessarily all of the intraday volume profile. A residual seasonality can survive, especially in the thin/heterogeneous overnight window (Asia / London / pre-market differ), so rvol / volume_zscore should not be read as fully seasonality-free.

No silent drift (D-12 / D-13)

• Existing base rates are byte-stable (D-12). All 19 pre-existing analyses' published rate / n / successes / ci_low / ci_high / min_n_pass are unchanged — the seven new features are purely additive standalone analyses that touch no existing compute or annotation column. This is enforced by scripts/build_conditioning_report.py, a drift gate that exits non-zero on any change to an existing analysis's base columns, run against the golden manifest before it is refreshed.
• Moved corrected verdicts are correct, not drift (D-13). Adding ~408 buckets does shift the bh_q_value / corrected_pass of some existing buckets, because apply_bh_correction builds one global BH family across every analysis × timeframe × instrument × direction in the discovery slice — never grouped per-analysis-key. A moved corrected verdict is the correct consequence of honestly accounting for more tested hypotheses (exactly what MHT-01 is for), not a base-rate drift: the base rates themselves do not move (proven by the drift gate); only the multiple-comparisons-corrected verdicts — which are the additive bh_* columns — do.

Corrected Null & Leakage Fixes (2026-09-29)

Quick task 260929-mkg (.planning/quick/260929-mkg-*). Everything here is additive: legacy columns (p_value, bh_*, corrected_pass, wf_verdict, verdict) are still written for provenance, but verdicts should read the geo_* columns.

1. Look-ahead leakage → confirm-then-enter. barrier_hit scores bars t+1..t+2, yet several conditioners were computed from those same bars. On NQ 1H (2020-01..OOS_START) a bullish-CISD mid0 FVG had a 0.1% chance of being stopped on t+1 (vs 33.6% without), a mid1 FVG had 0% timeouts, and a CISD-bar swing had a 0.00% stop-at-t+1 rate — the published 86–95% "edges" were the definition of the tag. Fix: barrier_hit_after(df, idx, row, ct, k) enters at close[t+k] with the same candle[0] target/stop, scores t+k+1..t+k+LOOKAHEAD, and excludes setups whose target or stop was touched before entry. cisd_fvg / fvg_size / Reading B use k=2, sssf_swing k=1; every bucket in an analysis shares its k. The expectancy study and forward-returns explorer anchor their FVG / structure cases the same way. Consequences worth knowing:

• A mid1 FVG requires bar t+2 to trade entirely beyond the CISD target, so it is never enterable at close[t+2] — mid1 buckets are empty by construction.
• Most mid0-FVG CISDs have already hit target by t+2; the few that are still enterable (NQ 1H bullish: 44 events in discovery) sit at their corridor baseline.
• cisd_fvg_interaction (hold state over the next 10 bars), smt_cisd "w/ SMT & survived/broke" (SMT broke on t+1..t+2) and candle1_followthrough *_inwindow are defined by the outcome window itself and cannot be fixed by moving entry: they carry diagnostic = True and are excluded from the corrected BH family and every geo verdict.
• tests/test_leakage_and_geo.py perturbs every bar after each tag's known-at bar on a random walk and asserts the tag does not change — a standing guard against new look-ahead.

2. Corridor-position null (with volatility strata and effective n). A bucket is tested against what its events' entry position in the stop→target corridor and volatility regime predict, not against 0.5. attach_geo_baseline() gives each CISD the hit rate of all same-slice CISDs with the same direction, scoring frame, ATR-regime tercile and 10%-corridor bin; a bucket's expected hits are the sum of those rates. atr_regime = ATR(14) as of t-1 ÷ its trailing 100-bar median. It is lagged one bar so the CISD candle's own range never feeds it, since that would co-vary with the body-size buckets. Terciles are cut within each baselined slice, and the warm-up NaN is its own stratum.

• Effective n. Events are clustered by CME session day (18:00 ET roll) on 15min/1H and by session week on 4H/Daily. The z-test uses a cluster-robust variance V = max(V_cluster, V_iid), which never credits negative within-cluster correlation. Manifests carry geo_clusters, geo_deff (design effect), geo_n_eff, geo_se, geo_z (clustered) and geo_z_iid (old independence z, provenance only). Every MIN_N gate on the geo path uses geo_n_eff.
• One test per bucket. Each bucket also gets a pooled NQ+ES row, in which same-day NQ and ES events share a cluster. The geo BH family and all geo verdicts use pooled rows only, so NQ and ES never count as two confirmations. Per-instrument rows keep clustered stats as description (geo_verdict = per-instrument). Pooled rows are excluded from the legacy 0.5-null family, so legacy numbers do not move.
• Equivalence, minimum effect 3 pp. geo_mde is the lift detectable with 80% power (two-sided 5%, uncorrected), and geo_equiv is a per-bucket TOST: the 90% CI of the lift lies inside ±3 pp. Pooled verdicts are above-baseline / below-baseline (BH-significant, geo_n_eff ≥ 50), within-3pp (not significant and TOST-equivalent), inconclusive (neither), below-n, diagnostic or no-baseline. inconclusive means the sample cannot tell, not that the effect is absent.
• What it still does not fix. Windows that straddle a cluster boundary are not merged. Few-cluster cells (Daily/4H weeks) have noisy variance estimates. The baseline includes the bucket's own events (indirect standardization), which shrinks lift toward zero. Nested and overlapping buckets (wick / wick_distance / combined, candle_size / size_cross, rvol / volume_zscore) are not independent findings. Over the whole population lift is 0 by construction, so basic always reads within-3pp.

3. Session tagging. Session was tagged at the bar label; 1H bars are labeled by their start, so the 09:00–10:00 bar (containing the open) was overnight and the 10:00–11:00 bar was rth_open. Tags are now taken at the bar midpoint (15min tags unchanged). Session cases/filters are intraday-only in the expectancy study and forward-returns explorer (review WR-01).

4. Data snapshot. The parquet files were replaced by a UTC-stamped vendor history (2010-06 → 2026-07-17). load_1m converts to ET and pins DATA_START..DATA_END = 2020-08-31..2025-11-21, which reproduces OOS_START and all four WALK_FORWARD_FOLDS exactly; the bars differ slightly from the lost original (characterization values re-pinned: intraday n ≤ 0.3%, rates ≤ 0.2 pp). 2025-11-22 → 2026-07-17 (HOLDOUT_END) was held out until the 2026-09-30 run below (build_validation.py --holdout).

Corrected results (regenerated 2026-09-30: discovery, walk-forward, fresh holdout)

Regenerated with SMT restored from github.com/thomas-quant/SMT @ 6300e0c. The original local copy was lost, so SMT numbers may differ slightly from the Phase 9 versions: characterization cells moved n −4..+7 and rates ≤ 0.7 pp, and that change cannot be split between the new data and the SMT version. Discovery and walk-forward verdicts were committed (66cfb21) before the holdout was run.

Discovery (pooled NQ+ES, 756 buckets):

Verdict	Buckets
above-baseline	37
below-baseline	25
within-3pp	146
inconclusive	283
below-n	88
diagnostic	104
no-baseline	73

The median design effect is 1.31 (IQR 1.17–1.44), so clustering removes ~25% of nominal n. It removes ~45% on Daily. The median MDE is 4.5 pp, which is why most untested-looking cells are inconclusive, not within-3pp. Of the 62 significant buckets, 52 are geo wf-robust across the 4 discovery folds. Those folds share the discovery data, so this is not independent confirmation.

Fresh holdout (2025-11-22 → 2026-07-17, never touched before this run): 52 of the 62 significant buckets have geo_n_eff ≥ 50 in the holdout.

• 51/52 keep the sign of their discovery lift, and 0 reverse significantly.
• 23/52 are individually significant (|z| ≥ 1.96) on 8 months of data.
• Lifts shrink: the median holdout/discovery ratio is 0.79 and the inverse-variance-weighted ratio is 0.85, the expected winner's-curse haircut.
• The old OOS slice (2024-04-30 → 2025-11-21, secondary check) agrees: 53/54 same sign, 31/54 significant, weighted ratio 0.82.
• The 51/52 is not 52 independent confirmations. The buckets nest and overlap heavily; the distinct families are roughly the eight in the table below.

Family (15min unless noted)	Discovery lift	Holdout lift (±95% CI)	Read
Candle body 1x–1.5x / >1.5x ATR	+4.7 to +6.6 pp	+3.5 to +7.6 (±3.7–5.4)	holds; 1H >1.5x +8.5 / +11.7 (±9)
Candle body <0.5x ATR	−1.8 / −2.4	−2.3 / −2.0 (±1.8)	holds
size_cross Big CISD / Small prev	+4.7 / +6.6	+5.1 / +5.5 (±3.5)	holds
Past-wick (wick, wick_distance, combined 1c)	+2.5 to +4.3	+2.2 to +3.5 (±2–3)	holds, smaller
post_cisd_context failed_gap_with / against	+3.0 / −3.6 (bull), +2.9 / −3.2 (bear)	+2.4 / −3.6, +1.2 / −2.5 (±2)	holds, bear side weaker
Low volume (rvol <0.7x, volume_zscore <−0.5σ)	−2.1 to −2.4	−0.6 to −3.3 (±2)	mixed: bearish holds, bullish fades
session rth_open	+3.6 / +3.8	+7.5 / +5.7 (±5.5)	survives the ATR-regime control; wide CI
cisd_fvg mid0 (enterable at close[t+2])	−7.9 to −11.2 below baseline	−3.2 (±12), n_eff ≈ 51	untestable in holdout

Within ±3 pp (equivalence, discovery). These buckets add less than 3 pp over corridor position and volatility (90% CI inside ±3 pp):

• sweep w/ sweep (15min and 1H): +0.7 to +1.5 pp.
• sssf_swing swing tags (15min): −0.7 to −0.1 pp.
• 15min smt_cisd w/ SMT: +1.5 pp (95% CI ±1.7).
• session rth / overnight.
• Most mc and effort_result cells.

On 4H/1H/Daily the same tags are inconclusive (MDE 5–16 pp), so nothing is claimed there.

Payoff of the same trade (scripts/build_barrier_payoff.py → output/barrier_payoff.{csv,md}). A rate lift is not a payoff lift: a higher hit rate usually means entry sits closer to the target, which makes the reward smaller in R. So the surviving buckets were re-scored on the trade the rate was measured on:

• The trade: entry at close[t+k], stop at the CISD candle's opposite extreme (−1R), target at its near extreme, 2 bars, otherwise marked at the close.
• What's reported: mean R, and the lift in mean R over the same corridor-bin × ATR-regime baseline, with cluster-robust 95% CIs.
• Costs are subtracted, not tested. The assumption is 1 tick of slippage per side plus ~$4 commission round trip, i.e. 0.70 pt NQ / 0.58 pt ES.

Results:

• The trade itself loses money before costs. All CISDs: mean R gross −0.12 (15min, 1H), −0.17 (4H), −0.20 (Daily). Net on 15min: −0.29 discovery, −0.23 holdout; costs are ~0.1R at a 6–11 pt median risk. Winners are small (entry near the target) and losers are the full candle.
• The surviving buckets make it less bad, not good. Their R lifts are real and replicate: 45/45 buckets with holdout n ≥ 100 keep the sign of their net R lift. For example, 15min >1.5x ATR body is +0.11 to +0.19 R over baseline, and past-wick is +0.05 to +0.10 R.
• But absolute expectancy stays ≤ 0.
  • In discovery, 55 of 62 buckets have a net-mean-R CI entirely below zero, and no bucket has a CI entirely above zero in either slice.
  • The best cells are roughly breakeven: 1H >1.5x ATR bearish −0.007 ± 0.031 R (holdout +0.009 ± 0.058), 1H far-past-wick bearish −0.005 ± 0.034 R, and 15min >1.5x ATR body −0.06 ± 0.02 R net in discovery (upper bound −0.04 R).
  • The Daily cells (+0.04 to +0.05 R) rest on n = 66–83 with ±0.07 CIs.
• Implication. Within this execution frame (CISD-candle target and stop, 2 bars), the data are consistent with a breakeven-at-best filter on a losing base trade, not with a tradable edge. The lifts may still matter as features for a different frame: another target/stop geometry, a longer horizon, or a sizing rule. That is a new hypothesis and has not been tested here.
• cisd_expectancy.{csv,md} (single barrier, open upside to 7 bars) was re-run with SMT. Its small positive gross R on NQ (+0.03 to +0.11 at 7 bars) has no CI, no costs, no corridor baseline, and runs over a strongly trending 2020–25 NQ. Read it as descriptive, not as an edge.

Caveats that travel with these numbers. (1) Nested and overlapping buckets are not independent. (2) The holdout is 8 months; Daily/4H cells are mostly below n there. (3) inconclusive is a statement about power, not about the market. (4) The baseline controls corridor position and ATR regime only. (5) Walk-forward folds share discovery data. (6) The per-bucket TOST is uncorrected for multiplicity.

Artifacts:

• output/validation_manifest_{discovery,walkforward,holdout,oos}.csv
• output/validation_findings_holdout.csv, output/post_cisd_verdict_holdout*.csv (primary)
• output/validation_findings.csv, output/post_cisd_verdict*.csv (old OOS, secondary)
• output/barrier_payoff.{csv,md} (payoff of the same barrier trade, discovery + holdout)
• Charts: the white tick on every bar is the corridor/ATR-regime baseline for that bucket.

Configuration

Edit constants at the top of cisd_analysis.py:

• LOOKAHEAD: Bars to check ahead (Default: 2).
• MAX_CONSEC: Max consecutive candles for Markov (Default: 3).
• TIMEFRAMES: Resampling rules (Default: Daily, 4H, 1H, 15min).
• FVG_HOLD_LOOKAHEAD: Bars used to judge whether a linked FVG held (Default: 10).
• SWEEP_TOLERANCE: Bars in the CISD-relative sweep window [t-4, t] (Default: 5).
• SWEEP_SWING_LOOKBACK: Lookback used to source prior swing points for sweeps (Default: 20).
• Swing SMT support depends on the local SMT package at /mnt/e/backup/code/Finance/Misc/SMT.

Requirements

• Python 3.12 (3.12.3 verified)
• Install pinned runtime: pip install -r requirements.txt
• Core libraries pinned in requirements.txt: pandas==3.0.2, numpy==2.4.4, matplotlib==3.10.8, pyarrow==23.0.1, pytest==9.0.2
• Plotly.js is loaded from CDN inside the generated output/forward_returns.html and is never pip-installed