18 KiB
18 KiB
P3.2 LABEL CONTRACT FORENSIC REPORT — SNIPERGOLD_ML
Date : 2026-08-22
Session : P3.2 (phase P3.2.1 — event extraction + MFE/MAE + time-to-outcome forensic)
Status : RESEARCH / DIAGNOSTIC ONLY — no production changes
Source : P2.6 corrected baseline + P3.1 (immutable checkpoints)
Artifacts : ml/p3/label_event_audit.py ; ml/p3/output/p3_2_label_events.json ;
ml/p3/output/p3_2_label_summary.json ; ml/p3/output/p3_2_deoverlap_check.json
0. PRODUCT QUESTION (defined before alternative labels — P3.2 §4)
SMC candidate setup occurs
-> would this setup be valid to trade?
-> did the expected outcome occur before invalidation?
Three concepts are distinguished:
directional prediction : future price direction (all bars)
setup validation : whether the EVENT setup produced the expected outcome
BEFORE invalidation <- WHAT MUST BE MEASURED
trade outcome : exploitation result (TP/SL, costs, slippage)
The existing label (24-bar / 0.75 ATR) measures DIRECTIONAL PREDICTION
(point-to-point drift close[i+24]-close[i]), NOT setup validation.
1. CHECKPOINT (P3.2 §2)
P2_SOURCE_SHA : 8d330343af688e2de2b2a1d12ce19a582714709a OK (HEAD == remote)
P2.6_SOURCE_SHA : 8d330343... (train_p26=e510ef3b, build_features_p2=4680f905)
FEATURE_CONTRACT_SHA : C44CC6F2... (canonical) | file 7b908b12... (discrepancy appendix, noted in P3.1)
P2.6_MODEL_SHA : 06df8452a112290ecb9bae1a3dbff6df492e084f8b872885d808bb0303ec6a70 OK
P2.6_DATASET_SHA : e85a08616191fb7ce4bbe79c1754228ddbcf54562034a322747501c7aabe124e OK (recomputed)
P3.1 reproducible : YES (deterministic script; baseline 0.5105/0.5063 still reproduced)
2. EVENT EXTRACTION (P3.2 §5 — D_EVENT)
Event families are built ONLY from existing semantics (FEATURE_CONTRACT v1.0). No new setup rules; no production meta-label gate.
| family | definition (existing semantics) | direction | n |
|---|---|---|---|
| E_BUY | f9==1 & f7>0 (CHoCH confirms bullish sweep) | long | 27096 |
| E_SELL | f9==1 & f7<0 (CHoCH confirms bearish sweep) | short | 25259 |
| E_BUY_STRICT | E_BUY & f18>=60 | long | 13143 |
| E_SELL_STRICT | E_SELL & f18>=60 | short | 9859 |
| E_EQH | f10==1 (equal-high swept) | short | 877 |
| E_EQL | f11==1 (equal-low swept) | long | 1617 |
Each event records: timestamp, direction, SMC state (from features), entry reference (close of the decision bar), and the outcome path (bars i+1..i+H).
3. ENTRY SEMANTICS (P3.2 §6)
Primary : A = close of the closed decision bar
(closed-bar causality; consistent with the existing label & runtime architecture)
Robustness (entry B = next-bar open):
median |close[i]-open[i+1]|/ATR = 0.008 ATR ; p95 = 0.055 ATR
-> the A vs B difference is NOT MATERIAL (< 0.06 ATR p95).
Decision: a single entry semantics (A) is used; B recorded as robustness.
4. MFE / MAE (P3.2 §7 L2)
Median excursions (ATR units), direction-aligned, per horizon:
| family | H=4 | H=8 | H=16 | H=24 |
|---|---|---|---|---|
| E_BUY MFE / MAE | 0.80 / 0.77 | 1.13 / 1.11 | 1.69 / 1.60 | 2.18 / 2.00 |
| E_SELL MFE / MAE | 0.78 / 0.78 | 1.12 / 1.15 | 1.65 / 1.68 | 2.10 / 2.09 |
| E_EQH MFE / MAE | 0.77 / 0.68 | 1.15 / 0.94 | 1.65 / 1.32 | 2.02 / 1.64 |
| E_EQL MFE / MAE | 0.72 / 0.78 | 1.03 / 1.06 | 1.48 / 1.49 | 1.79 / 1.73 |
E_BUY / E_SELL : MFE ~= MAE at ALL horizons (symmetric).
-> no favorable payoff asymmetry on the f9-confirm setup.
E_EQH (all) : MFE > MAE (1.65 vs 1.32 at H16) — THE ONLY asymmetry.
BUT: after de-overlap (see §8) the asymmetry DISAPPEARS (overlap artifact).
E_EQL : symmetric up to MAE > MFE (slightly unfavorable).
5. TP-BEFORE-SL (P3.2 §7 L3 + §9 asymmetric)
L3 = WIN (TP before SL) / LOSS / UNRESOLVED / AMBIGUOUS. Semantic combos: 1.0/0.5, 1.5/0.75, 2.0/1.0 (R:R 2:1), 1.0/1.0 (symmetric).
| family | H | combo | P_WIN | P_LOSS | P_UNRES | P_AMB | med tTP | med tSL |
|---|---|---|---|---|---|---|---|---|
| E_BUY | 16 | 1.0/0.5 | 0.323 | 0.638 | 0.012 | 0.027 | 3 | 2 |
| E_BUY | 16 | 1.5/0.75 | 0.310 | 0.629 | 0.052 | 0.008 | 5 | 3 |
| E_BUY | 16 | 2.0/1.0 | 0.282 | 0.596 | 0.116 | 0.005 | 7 | 3 |
| E_BUY | 24 | 1.0/1.0 | 0.488 | 0.478 | 0.022 | 0.012 | 4 | 4 |
| E_SELL | 16 | 1.5/0.75 | 0.310 | 0.633 | 0.048 | 0.009 | 5 | 3 |
| E_SELL | 24 | 1.0/1.0 | 0.487 | 0.483 | 0.018 | 0.012 | 4 | 4 |
| E_EQH | 24 | 1.0/1.0 | 0.525 | 0.436 | 0.039 | 0.001 | 4 | 5 |
| E_EQH | 16 | 1.5/0.75 | 0.326 | 0.577 | 0.092 | 0.005 | 5 | 3 |
| E_EQL | 24 | 1.0/1.0 | 0.466 | 0.495 | 0.032 | 0.007 | 4 | 4 |
1. E_BUY/E_SELL: LOSS > WIN for ALL R:R>1 combos and ALL horizons
(consistent pattern: LOSS ~2x WIN for 1.5/0.75).
2. tSL < tTP consistently -> SL is touched FIRST:
the f9-confirm setup is a LATE ENTRY that often hits a pullback before the target.
3. 1.0/1.0 H24 is near 50/50 (E_BUY 0.488/0.478) -> without distance asymmetry,
there is no edge.
4. E_EQH all points to WIN>LOSS (0.525/0.436) — BUT it is an overlap artifact (§8).
5. Conclusion: NO event family has a TP-before-SL asymmetry that survives
de-overlap. Asymmetric TP/SL does not help because the problem is in the
timing/late-entry, not the distance.
6. TIME-TO-OUTCOME (P3.2 §7 L4 + §8)
median time-to-TP : 2-8 bars (increases with TP distance & horizon)
median time-to-SL : 1-5 bars (always <= tTP)
median time-to-MFE: 2-12 bars ; time-to-MAE: 2-11 bars (H4->H24)
Setups empirically "live" very briefly:
most outcomes (TP or SL) occur within 1-8 bars after the event.
The 24-bar horizon FAR exceeds the setup lifetime; the 24-bar label captures
long-term drift UNRELATED to the event.
7. HORIZON ANALYSIS (P3.2 §8)
| H | E_BUY 1.5/0.75 W/L/U | E_SELL 1.5/0.75 W/L/U | E_EQH 1.5/0.75 W/L/U | E_EQL 1.5/0.75 W/L/U |
|---|---|---|---|---|
| 4 | 0.190/0.487/0.315 | 0.195/0.490/0.308 | 0.203/0.433/0.359 | 0.162/0.489/0.340 |
| 8 | 0.265/0.581/0.146 | 0.266/0.587/0.139 | 0.290/0.520/0.186 | 0.231/0.584/0.176 |
| 16 | 0.310/0.629/0.052 | 0.310/0.633/0.048 | 0.326/0.577/0.092 | 0.280/0.631/0.080 |
| 24 | 0.322/0.640/0.029 | 0.321/0.644/0.026 | 0.344/0.603/0.048 | 0.291/0.644/0.056 |
- Censoring (UNRESOLVED) drops fast: H4 30-36% -> H8 14-19% -> H16 5-9% -> H24 2-6%.
- P_WIN is relatively stable after H8 (asymptote ~0.31-0.34); P_LOSS keeps rising.
- No horizon where the setup shows WIN > LOSS (de-overlapped).
- The 8-16 bar horizon captures most outcomes with low censoring;
the 24-bar horizon adds nothing but additional SL.
8. EVENT OVERLAP AUDIT (P3.2 §12)
E_BUY : median gap 1 bar | P(next event within 16 bars) = 98.5% | cluster med 46, max 388
P_isolated = 0.05% (14 of 27096!)
E_SELL : median gap 1 bar | P(next within 16) = 98.5% | cluster med 46, max 473
P_isolated = 0.08%
E_EQH : median gap 1 bar (bimodal; mean 202) | P(next within 16) = 98.1% | cluster 17-384
E_EQL : median gap 1 bar | P(next within 16) = 98.9% | cluster 42, max 504
MASSIVE overlap. f9-confirm is a state living in long runs (median cluster 46 bars): consecutive f9 events share the outcome window -> repeated counting. De-overlap (lead event per cluster, separation > 16 bars):
| family | all -> lead | 1.5/0.75 H16 W/L | 1.0/1.0 H24 W/L |
|---|---|---|---|
| E_BUY | 27096 -> 1822 | 0.321 / 0.621 | 0.499 / 0.465 |
| E_SELL | 25259 -> 1680 | 0.315 / 0.634 | 0.485 / 0.484 |
| E_EQH | 877 -> 62 | 0.258 / 0.661 | 0.435 / 0.500 |
| E_EQL | 1617 -> 109 | 0.294 / 0.633 | 0.486 / 0.468 |
AFTER DE-OVERLAP:
- The E_EQH asymmetry (0.525/0.436) DISAPPEARS -> 0.435/0.500. OVERLAP ARTIFACT.
- All families: LOSS >= WIN (except E_BUY 1.0/1.0 H24 0.499/0.465 — 50/50).
- MFE ~= MAE for all families.
CONCLUSION: the event-based formulation WITH existing semantics remains highly
serial-dependent; event-based labels MUST be de-overlapped (lead event per cluster)
before interpretation. And after de-overlap, there is no edge.
9. INVALIDATION SEMANTICS (P3.2 §10)
Uses existing semantics: opposite structure break = f8 (choch_dir) or f5 (chart_bias) flipping sign against the event direction.
| family | H=16 P(opp break within H) | P(opp break BEFORE TP) |
|---|---|---|
| E_BUY | 0.538 | 0.474 |
| E_SELL | 0.594 | 0.526 |
| E_EQH | 0.881 | 0.828 |
| E_EQL | 0.933 | 0.887 |
- f9-confirm BUY/SELL: the opposite structure occurs BEFORE TP on ~47-53% of
events -> structure often cancels the setup.
- EQH/EQL: 83-89% (the EQH/EQL bias direction is often opposed by structure —
consistent with E_EQH/E_EQL having no favorable asymmetry).
- Invalidation (opposite break) is a relevant semantic candidate for a
"failed before target" label, BUT its frequency is very high for EQH/EQL.
10. CENSORING / AMBIGUITY (P3.2 §11)
AMBIGUOUS (TP & SL on the same bar): 0.5% - 2.7% across all families/combos -> LOW.
(no need to pick a favorable ordering; AMBIGUOUS excluded/recorded.)
Censoring (UNRESOLVED):
H4 : 12-51% (high) -> short horizons need survival-style treatment.
H8 : 14-19% (1.5/0.75)
H16 : 5-9% (low) -> binary TP-before-SL is VIABLE.
H24 : 2-6% (minimal)
Conclusion: at H16+, binary WIN/LOSS/UNRESOLVED is viable; UNRESOLVED must NOT be
forced to LOSS (real censoring, not failure).
11. FEATURE INFORMATION UNDER CANDIDATE LABELS (P3.2 §13)
The frozen P2 features (19) are evaluated univariately against the L3 label (WIN vs LOSS) on the event subset, WITHOUT retraining. P3.1 baseline: max |AUC-0.5| = 0.014.
| configuration | E_BUY max|dev| (n) | E_SELL max|dev| (n) |
|---|---|---|
| H8 TP1.0/SL0.5 | 0.0210 (25276) | 0.0108 (23645) |
| H8 TP2.0/SL1.0 | 0.0361 (19617) | 0.0225 (18599) |
| H16 TP2.0/SL1.0 | 0.0309 (23809) | 0.0163 (22368) |
| H24 TP2.0/SL1.0 | 0.0310 (25110) | 0.0169 (23559) |
| MI max | 0.0005 (E_BUY) / 0.0010 (E_SELL H24) |
- Information content RISES ~2.5x from baseline (0.014 -> 0.036) when the label
is aligned with the event (E_BUY, farther TP 2.0 ATR).
- E_BUY is consistently > E_SELL -> the long side is slightly more "predictable"
by the frozen features against the event-aligned label.
- However, 0.036 is STILL below the practical threshold (~0.05) and not yet
de-overlapped (part could be overlap artifact).
- Conclusion: the event-aligned label starts to "activate" feature information,
but it is not yet material. This is partial evidence for "features informative
against the right label" vs "features are truly uninformative" — not yet
conclusive (needs de-overlap + final label).
12. LABEL QUALITY SCORECARD (P3.2 §14)
| Criterion | L1 directional | L2 MFE/MAE | L3 TP-before-SL | L4 time-to-outcome |
|---|---|---|---|---|
| Causal | PASS | PASS | PASS | PASS |
| SMC-aligned | FAIL | PASS | PASS | PASS |
| Execution-realistic | PASS | PASS | PASS | PASS |
| Event-specific | FAIL | PASS | PASS | PASS |
| Low overlap | FAIL (all bars) | PARTIAL (de-overlap) | PARTIAL (de-overlap) | PARTIAL (de-overlap) |
| Low ambiguity | PASS | PASS | PASS (1-3%) | PASS |
| Low censoring | PASS | PASS | PASS (H16+) | PARTIAL (survival) |
| Economically interpretable | FAIL (drift) | PASS (risk) | PASS (W/L) | PASS |
| Suitable for meta-labeling | FAIL | FAIL (not binary) | PASS | FAIL (survival) |
L3 (TP-before-SL, event-based, de-overlapped, H16+) = the only family meeting the
meta-labeling requirements: causal, SMC-aligned, event-specific, low ambiguity,
low censoring, binary, economic.
L2/L4 are important for understanding setup characteristics, not as classification targets.
13. ROOT-CAUSE CLASSIFICATION (P3.2 §21/§22-M)
B — LABEL IS MISALIGNED : STRONGLY SUPPORTED
- t_hit ~2 bars & typical MFE 3.8 ATR (P3.1): small threshold vs noise path
- L1 (drift) vs L3 (TP-before-SL) give DIFFERENT profiles: L1 ~50/50,
L3 LOSS>WIN for f9-confirm -> the 24-bar label measures the wrong thing
- setups live 1-8 bars, not 24 bars (time-to-outcome)
C — LABEL FAMILY REQUIRED : PARTIALLY SUPPORTED
- H4 censoring high -> survival/MFE-MAE needed for short horizons
- for the 8-16 horizon, binary L3 is viable; L2/L4 as descriptive complements
A — FEATURE/SETUP FAILURE : REMAINS VALID (P3.1 finding, confirmed in P3.2)
- after de-overlap NO event family has a payoff asymmetry
- MFE ~= MAE; LOSS >= WIN on all existing setups
- label redesign ALONE will not activate an edge: the existing setup
semantics (f9-confirm, EQH/EQL as features) produce no asymmetry
Official P3.2 classification: B — LABEL IS MISALIGNED (for setup-validation
purposes), with the strong note that label redesign is necessary BUT NOT
SUFFICIENT: the existing setup/feature semantics also produce no payoff
asymmetry after de-overlap (confirmation of finding A from P3.1).
14. CANDIDATE LABEL CONTRACT (P3.2 §21 — PROPOSED, not deployed)
NAME : SGML_L3_EVENT (candidate, not yet frozen)
UNIT : only SMC EVENT bars (D_EVENT), not all bars
ENTRY : close of the decision bar (A); next-bar open (B) almost identical (0.008 ATR)
OUTCOME : TP-before-SL
TP = 1.5 x ATR(entry bar) ; SL = 0.75 x ATR(entry bar) [semantic 2:1]
(1.0/1.0 as robustness; 2.0/1.0 for far targets)
HORIZON : 16 bars (4 hours) primary ; 8 bars (2 hours) short ; 24 bars (6 hours) long
CLASSES : WIN / LOSS / UNRESOLVED / AMBIGUOUS
(UNRESOLVED & AMBIGUOUS are NOT forced to LOSS)
DE-OVERLAP: lead event per cluster (gap > 16 bars) OR first event of a cluster;
isolated/overlapping strata reported
INVALIDATION: opposite structure break (f8/f5 flip) recorded as a secondary
metric, not a replacement for SL
PROVENANCE: dataset hash, contract hash, script hash, purged split, seed
IMPORTANT NOTE: this label is a CANDIDATE CONTRACT for correctly measuring setup validation — NOT guaranteed to produce an edge. P3.2 evidence shows existing setups most likely REMAIN edgeless under this label (MFE≈MAE, LOSS≥WIN). The value of this label is a CORRECT measurement, not an edge creator.
15. RECOMMENDATION (P3.2 §22-O)
1. The 24-bar/0.75 ATR label is NOT used as a setup-validation measure.
(It remains valid as a directional-drift measure; not for SMC setups.)
2. Propose SGML_L3_EVENT (section 14) as a CANDIDATE LABEL CONTRACT.
Review architecture/research first; DO NOT retrain before explicit approval.
3. Deeper findings for future research (not this session):
- existing setup semantics (f9-confirm & EQH/EQL as features) have NO payoff
asymmetry after de-overlap -> the problem is at the SETUP/FEATURE
ARCHITECTURE level, not only the label (P3.1 A remains valid).
- re-examine the true "candidate setup" definition for SMC (order block,
breaker, FVG, displacement) as a feature-architecture revision — outside
the P3.2 mandate (no hot-fix).
4. LSTM / regime / meta-label REMAIN UNTESTED until the label contract is
approved and the feature re-audit against the new label is complete.
16. SUCCESS CRITERIA (P3.2 §20)
| # | Criterion | Answer |
|---|---|---|
| 1 | What does "setup succeeded" mean? | The expected outcome (TP) occurs BEFORE invalidation (SL / opposite break), within the setup-lifetime horizon |
| 2 | Does 24-bar/0.75 ATR measure that? | NO — it measures directional drift; threshold small vs noise (MFE 3.8); horizon >> setup lifetime (1-8 bars) |
| 3 | How long do setups live? | 1-8 bars (median tTP 2-8, tSL 1-5); low censoring at H16+ |
| 4 | Is TP-before-SL more representative? | YES (L3) — causal, SMC-aligned, event-specific; ambiguity 1-3%; low censoring at H16+ |
| 5 | How are MFE/MAE? | Symmetric for all families after de-overlap (MFE ≈ MAE) |
| 6 | How large is event overlap? | MASSIVE: median gap 1 bar; P(next within 16) 98%; cluster med 46 bars; isolated 0.05% |
| 7 | Censoring/ambiguity? | Ambiguity 1-3% (low); censoring H4 12-51%, H16 5-9%, H24 2-6% |
| 8 | Does feature information appear with an event-aligned label? | PARTIAL: max |AUC-0.5| 0.014 -> 0.036 (2.5x), E_BUY>E_SELL, still below the practical threshold & not de-overlapped |
17. EVIDENCE CLASSIFICATION
TRUSTED : all P3.2.1 tables (deterministic script, provenance recorded,
dataset hash matches, de-overlap check recorded)
massive-overlap findings (98%+), MFE≈MAE, LOSS≥WIN after de-overlap
INCONCLUSIVE : feature information under the new label (0.036, not yet de-overlapped) —
whether features are truly informative against the final label is unresolved
INVALID : the favorable E_EQH asymmetry in all-events (overlap artifact,
gone after de-overlap)
SUPERSEDED : none (P2.6/P3.1 baselines remain valid for their purpose)
Appendix — P3.2 Artifacts & Provenance
Script : ml/p3/label_event_audit.py (SHA recorded in summary)
Output : ml/p3/output/p3_2_label_events.json (per-event, H16, combo 1.5/0.75)
ml/p3/output/p3_2_label_summary.json (full summary)
ml/p3/output/p3_2_deoverlap_check.json (lead-event check)
Data : Files\AlgoForge\Data\XAUUSD_*.npz (P3.1 hash)
P2.6 dataset : e85a0861...
Frozen model : 06df8452...
Entry semantics : close of the decision bar (A); B checked (0.008 ATR med)
Horizons : 4/8/16/24 M15 bars (1h/2h/4h/6h)
ATR : rolling 14-bar M15 (contract)
Random seed : not used (deterministic)
Closing note: P3.2 completed without production changes. FEATURE_CONTRACT, runtime semantics, and the baseline model were NOT touched. The candidate label (SGML_L3_EVENT) is a research proposal, NOT a deployment. The finding that existing setups have no payoff asymmetry after de-overlap is a forensic result that should direct future research to a SETUP/FEATURE architecture review, not to more complex models.