Aliwkarimi98/SniperGold_ML
Watch
1
0
Fork
You've already forked SniperGold_ML
 1
forked from chiki2bum2/SniperGold_ML
Code Pull requests Activity
SniperGold_ML/docs/P3_S21_1_FEATURE_CONTRIBUTION_AUDIT.md

479 lines
23 KiB
Markdown
RawPermalinkNormal ViewHistory

docs: record P3-S21.1 feature contribution results
2026-08-25 16:46:36 +07:00
# P3-S21.1 FEATURE CONTRIBUTION AUDIT — FINAL REPORT
```text
Date : 2026-08-25
Session : P3-S21.1 — Feature Contribution Audit (frozen Logistic, P3-S20 folds)
Status : COMPLETE
Verdict : D — MIXED CONTRIBUTION (ZONE-leading, CONTEXT secondary; both
sign-stable across the three pre-registered folds; weak ranking
edge with redundant-by-construction feature structure)
Question : WHICH FROZEN FEATURES CONSISTENTLY CONTRIBUTE TO THE WEAK LOGISTIC
SIGNAL OBSERVED IN P3-S20, AND IS THAT CONTRIBUTION CONSISTENT
ENOUGH TO DESERVE FURTHER INVESTIGATION?
Scope : DESCRIPTIVE COEFFICIENT AUDIT ONLY. NO calibration, NO new model,
NO ablation, NO feature selection/removal, NO TP/SL/horizon/label
change, NO Candidate Setup/MQL5/FEATURE_CONTRACT change, NO
external data, NO hyperparameter tuning.
Human verif: REMAINS CANCELLED (historical only).
```
---
## A. Latest SESSION_HANDOVER used
```text
docs/SESSION_HANDOVER_2026-08-24_P3_S20_WALK_FORWARD.md
(newest authoritative handover, committed at HEAD 7c9602c; read completely
FIRST, before any diagnostic — session-handover-first rule).
Reconciles with Git (verified before any work):
local HEAD == origin/main == 7c9602ca4964f42f4ae8b7ac1ea14ab5e49e9176
branch main, working tree CLEAN,
origin = https://forge.mql5.io/chiki2bum2/SniperGold_ML.git
Handover confirms:
P3-S.20 = COMPLETE
verdict = A — STABLE WEAK SIGNAL (pre-registered rule output)
P3-S.21 = NOT STARTED (owner decision required)
next = feature-mechanism + calibration diagnostics (this session is the
feature-mechanism diagnostic: P3-S21.1, explicitly authorized).
Required documents (all read completely):
docs/P3_S20_WALK_FORWARD_BASELINE.md (A verdict, folds, quality gates)
docs/P3_S19_DATASET_FEATURE_AUDIT.md (audit PASSED; 686=594+92)
docs/P3_S18_BASELINE_ML_REPORT.md (INCONCLUSIVE / DATA TOO SMALL)
docs/P3_S18A_LABEL_CONTRACT_REVIEW.md (label contract APPROVED AS V1)
docs/P3_S16_LABEL_CONTRACT.md (v1, immutable)
docs/P3_S16_SETUP_DATASET_CONTRACT_v1.md (v1, immutable)
```
## B. Checkpoint
```text
Starting checkpoint (verified): 7c9602ca4964f42f4ae8b7ac1ea14ab5e49e9176
local HEAD == origin/main, branch main, working tree CLEAN,
origin = forge.mql5.io/chiki2bum2/SniperGold_ML.git
no stashes; reflog top = P3-S20 docs commit.
Final Forge HEAD : <P3_S21_1_FINAL_SHA recorded after push>
```
## C. Frozen Logistic configuration
```text
The audit refits the EXACT P3-S20 frozen configuration per fold and VERIFIES
byte-level reproduction before using any coefficient:
Model : LogisticRegression(C=1.0, L2 via default lbfgs solver,
max_iter=5000, random_state=42, fit_intercept=True,
class_weight=None) — identical to walk_forward.py.
Preprocess : StandardScaler(z=(x-mean)/std) fitted on THAT fold's TRAINING
rows only; applied unchanged to OOS.
Folds : expanding-window, chronological (P3-S20 pre-registered):
F1 train[0:300] OOS[300:395]; F2 train[0:395] OOS[395:490];
F3 train[0:490] OOS[490:571]; purge gaps 555/472/321 bars.
Coefficients reported: std_coef = coefficient on the STANDARDIZED input
(the quantity the frozen model actually used); raw_coef =
std_coef / scaler.scale_ (back-transformed to original units).
REPRODUCTION (hard gate, section 7):
OOS WIN-probabilities recomputed from a fresh frozen refit match the
committed ml/p3/baseline/output/p3_s20_oos_predictions.csv EXACTLY:
max_abs_prob_diff = 0.00e+00 (271/271 rows, fold/setup_id/outcome match)
pooled OOS ROC-AUC recomputed = 0.5791603786527645 == committed.
-> exact reproduction established; the coefficients below ARE the P3-S20
fitted decision function's coefficients. STOP condition not triggered.
```
## D. Feature inventory
```text
Frozen 12-feature schema (feature_sha16 = 0414e401522ea4e2, UNCHANGED).
Statistics below are over the 571 binary WIN/LOSS lead rows (the P3-S20
population); missing rate is 0.0 for all features. Full table:
ml/p3/baseline/output/p3_s21_1_feature_inventory.csv
feature family mean median std unique const
direction CONTEXT -0.0088 -1.0 1.000 2 no
h4_gate CONTEXT -0.0088 -1.0 1.000 2 no
m30_gate CONTEXT -0.0088 -1.0 1.000 2 no
sweep_age_bars LIQ/STRUCT 20.37 19.0 11.316 41 no
choch_age_bars LIQ/STRUCT 4.36 0.0 7.295 34 no
choch_latency_bars LIQ/STRUCT 16.01 14.0 10.140 41 no
zone_type_code ZONE 1.0 1.0 0.0 1 YES
zone_age_bars ZONE 8.99 3.0 30.238 44 no
zone_width_atr ZONE 1.194 1.047 0.684 571 no
price_in_zone_offset ZONE 0.4066 0.493 2.308 570 no
dist_to_zone_center_atr ZONE -0.0934 -0.0068 2.308 571 no
atr_at_entry SCALE 3.119 2.12 2.971 568 no
Semantic families (predefined, per section 9, from feature definitions):
CONTEXT = direction, h4_gate, m30_gate
LIQUIDITY_STRUCTURE= sweep_age_bars, choch_age_bars, choch_latency_bars
ZONE = zone_type_code, zone_age_bars, zone_width_atr,
price_in_zone_offset, dist_to_zone_center_atr
SCALE = atr_at_entry
ENTRY_MOMENTUM = EMPTY (no M15 displacement/entry-condition column in
the frozen 12)
MICRO = EMPTY (no M3 confirmation column in the frozen 12)
```
## E. Feature coefficients
```text
Standardized coefficients (on the standardized input the frozen model used),
per fold. Full table: ml/p3/baseline/output/p3_s21_1_fold_coefficients.csv
and ml/p3/baseline/output/p3_s21_1_feature_coefficients.csv.
feature std F1 std F2 std F3 mean|std| pattern
direction -0.0945 -0.0688 -0.0824 0.0819 all -
h4_gate -0.0945 -0.0688 -0.0824 0.0819 all -
m30_gate -0.0945 -0.0688 -0.0824 0.0819 all -
sweep_age_bars -0.0089 +0.0303 +0.0323 0.0238 mixed
choch_age_bars -0.0313 -0.0686 -0.0503 0.0501 all -
choch_latency_bars +0.0108 +0.0806 +0.0710 0.0541 all +
zone_type_code 0.0 0.0 0.0 0.0 zero
zone_age_bars -0.0217 -0.0077 +0.0160 0.0151 mixed
zone_width_atr -0.0786 -0.0939 -0.0642 0.0789 all -
price_in_zone_offset +0.1186 +0.1251 +0.1705 0.1380 all +
dist_to_zone_center_atr +0.1186 +0.1251 +0.1705 0.1380 all +
atr_at_entry +0.0509 +0.0490 +0.1398 0.0799 all +
Top-3 by mean |std coef|:
1. price_in_zone_offset / dist_to_zone_center_atr (0.1380 each)
2. direction / h4_gate / m30_gate (0.0819 each)
3. zone_width_atr (0.0789)
Wording: these are contributions to the FITTED LINEAR DECISION FUNCTION of
the frozen logistic, NOT causal importance, NOT independent predictive power,
NOT robustness, NOT production usefulness.
```
## F. Fold-level coefficient stability
```text
feature F1 rank F2 rank F3 rank sign pattern class
direction/h4/m30 3 5 4 all negative STABLE
sweep_age_bars 11 10 10 mixed INCONCLUSIVE
choch_age_bars 8 8 9 all negative STABLE
choch_latency_bars 10 4 7 all positive VARIABLE
zone_type_code 12 12 12 zero INCONCLUSIVE
zone_age_bars 9 11 11 mixed INCONCLUSIVE
zone_width_atr 6 3 8 all negative STABLE
price_in_zone_offset 1 1 1 all positive STABLE
dist_to_zone_center_atr 2 2 2 all positive STABLE
atr_at_entry 7 9 3 all positive VARIABLE
Sign consistency : direction/h4/m30, choch_age, zone_width,
price_in_zone_offset, dist_to_zone_center_atr are
SAME-SIGN in all 3 folds.
Magnitude spread : atr_at_entry jumps in F3 (0.050 -> 0.140) and
choch_latency grows from F1 (0.011) to F2/F3
(0.08/0.07) -> classified VARIABLE (same sign, spread
> 2.5x).
Fold disagreement: sweep_age (F1 - / F2,F3 +) and zone_age (F1,F2 - /
F3 +) FLIP sign -> INCONCLUSIVE (no consistent sign).
No feature is selected as "important" because it is strongest in one fold;
ranks 1-2 (zone offset pair) and the direction triple are the only
features in the top-4 in >= 2 folds with a stable sign.
```
## G. Group-level contribution
```text
Aggregation: combined = mean over folds of sum(|std coef|) of the group's
members (deterministic, documented). Sign stability = dominant member's sign
stable across folds. Full table:
ml/p3/baseline/output/p3_s21_1_group_contributions.csv
group n combined_mean F1 F2 F3 sign_stable dominant
CONTEXT 3 0.2457 0.283 0.206 0.247 True direction
LIQUIDITY_STRUCTURE 3 0.1280 0.051 0.179 0.154 True choch_latency
ZONE 5 0.3701 0.337 0.352 0.421 True price_in_zone_offset
SCALE 1 0.0799 0.051 0.049 0.140 True atr_at_entry
ENTRY_MOMENTUM 0 0.0 - - - - -
MICRO 0 0.0 - - - - -
ZONE carries the largest combined contribution (0.370), CONTEXT second
(0.246). CAUTION: CONTEXT mass is inflated ~3x by the identical
direction/h4/m30 columns and ZONE mass is inflated ~2x by the affine
price_in_zone_offset / dist_to_zone_center_atr pair (see section I) — the
unique-dimension ordering is: (1) ZONE entry-position/offset, (2) CONTEXT
direction/gate, (3) ZONE width, (4) ATR scale, (5) CHoCH timing.
Aggregation is used for DESCRIPTION only; nothing is removed.
```
## H. Feature distribution by fold
```text
Per-fold TRAINING-row distributions: ml/p3/baseline/output/p3_s21_1_fold_feature_stats.csv
(mean/median/std/min/max/q05/q25/q50/q75/q95; missing=0 everywhere).
Key comparisons (train rows):
price_in_zone_offset : F1 mean 0.422 | F2 0.415 | F3 0.372; medians
0.465/0.436/0.448 — NO material shift; F2's stronger
result is NOT explained by a different offset
distribution (mean/std nearly identical).
zone_width_atr : F1 1.184 | F2 1.191 | F3 1.195 (stable, small).
atr_at_entry : F1 2.026 | F2 2.011 | F3 2.381, max jumps to 19.97
in F3 (vs 8.72 before) — the F3 ATR-coefficient jump
(+0.140) coincides with a wider/right-skewed ATR
regime in the F3 training window.
direction/h4/m30 : mean -0.02/-0.033/-0.029 (short-favoring throughout).
choch_age_bars : medians 0 in all folds; mean ~4.2-4.3 (stable).
Conclusion: F2's OOS improvement is NOT attributable to a distinct feature
distribution in the offset/gate/width features; it appears period-specific
(temporal heterogeneity), consistent with P3-S20's own fold analysis. No
feature adaptation, no per-fold renormalization was performed (the frozen
train-only scaler was respected).
```
## I. H4/M30/direction collinearity
```text
VERIFIED (matches P3-S19 result, exactly):
h4_gate == m30_gate == direction on every one of the 571 binary rows.
unique combinations: only (-1,-1,-1) and (+1,+1,+1).
direction/h4/m30 have identical standardized columns -> identical
coefficients by construction.
NEW STRUCTURAL FINDING (affine identities in the verified population):
dist_to_zone_center_atr == price_in_zone_offset - 0.5 on every row
(the zone-center offset is the in-zone offset shifted by half a zone
width; the two features carry the SAME information after centering).
choch_latency_bars == sweep_age_bars - choch_age_bars on every row
(CHoCH latency is defined as sweep age minus CHoCH age — a functional
identity, not a new independent dimension).
zone_type_code is CONSTANT (all in-scope setups share one zone type).
Interpretation: these are EXACT structural/affine identities from the
verified canonical chain (F3 hard gates must be direction-compatible;
zone-center offset = in-zone offset - 0.5 width; latency = age difference).
They explain the identical coefficients; they do NOT make any feature leaky
or invalid, and NO feature is removed. The effective independent dimension
count of the 12 columns is ~8 (12 minus 2 exact duplicates minus 1 constant
minus 1 functional combination), NOT a defect.
```
## J. Zone geometry / ATR mechanism
```text
Exact definitions (from prepare_dataset.py, all causally available at entry
= creation-bar close):
zone_width_atr : (zone_top - zone_bot) / ATR(entry); 0 if degenerate.
units = ATR-ratio (width in ATR multiples).
price_in_zone_offset : (entry_price - zone_bot) / zone_width in [0,1];
entry position within the zone.
dist_to_zone_center_atr : (entry_price - zone_midpoint) / zone_width (signed);
NOTE the divisor is ZONE WIDTH, not ATR, despite
the column name — documented exactly here.
atr_at_entry : rolling 14-bar M15 ATR ending at the entry bar.
Availability at entry: ALL zone/ATR values are computed from closed bars and
the zone bounds as-of the creation bar — the entry bar itself is closed, so
every value is known before any future bar. ATR(entry) also scales the TP/SL
label; this is CONTEXTUAL SCALE dependence, NOT leakage: the value is known
at entry and does not reveal the outcome (P3-S19 section J/K; S18A N).
The model's persistent positive weight on entry position within the zone
(+0.119/+0.125/+0.170 across folds) and persistent negative weight on zone
width (-0.079/-0.094/-0.064) are consistent with a zone-geometry mechanism;
this audit does not claim causality.
```
## K. Mechanistic interpretation
```text
Persistent contributors and candidate mechanisms (language: "consistent
with" / "associated with", NOT "causes"):
price_in_zone_offset / dist_to_zone_center_atr (+ stable, ranks 1-2 all
folds): entry position relative to the zone. Candidate mechanism: setups
entering deeper into the zone (larger in-zone offset, i.e., nearer the
far edge of the zone in the setup direction) are associated with higher
WIN log-odds. This is the single most persistent contributor.
direction / h4_gate / m30_gate (- stable, one dimension): short-favoring
context. Candidate mechanism: in this historical XAUUSD population,
short-context setups are weakly associated with higher WIN log-odds
(consistent with P3-S18A's descriptive short/long counts). Because the
gates are equal to direction by construction, this is one context
dimension, not three.
zone_width_atr (- stable): wider zones (relative to ATR) are associated
with LOWER WIN log-odds. Candidate mechanism: a wider zone implies entry
farther from a tight invalidation anchor (0.75 ATR SL) relative to the
1.5 ATR TP — geometrically consistent with a width-vs-outcome distance
relationship. Descriptive only.
atr_at_entry (+ but VARIABLE, F3 jump): higher ATR regime associated with
higher WIN log-odds in F3; not stable enough for a confident mechanism.
choch_age_bars (- stable) / choch_latency_bars (+ variable): fresher CHoCH
(smaller age) weakly associated with higher WIN log-odds; latency's sign
is stable but its magnitude varies across folds.
INCONCLUSIVE: sweep_age_bars, zone_age_bars (sign flips), zone_type_code
(constant, zero coefficient).
No causal claim is made; these are candidate mechanisms of the frozen fitted
decision function only.
```
## L. Reproducibility
```text
The audit was executed multiple times with identical results:
- all 5 CSVs BYTE-IDENTICAL across runs (verified by fc /b):
p3_s21_1_feature_inventory.csv
p3_s21_1_feature_coefficients.csv
p3_s21_1_fold_coefficients.csv
p3_s21_1_group_contributions.csv
p3_s21_1_fold_feature_stats.csv
- p3_s21_1_summary.json identical except generated_utc (git_commit same
HEAD) — verified programmatically: only 'generated_utc' differed.
- deterministic tests S21.1-T01..T07 ALL PASS (7/7):
T01 feature schema identical to P3-S20
T02 label contract unchanged
T03 P3-S20 fold boundaries unchanged
T04 frozen Logistic configuration reproduced
T05 coefficient calculation deterministic
T06 group assignment deterministic
T07 repeated execution identical
- frozen parity-absence guards (P3-S.4/P3-S.5) scan CLEAN across ml/**/*.py
(0 hits) after adding the new scripts.
Recorded provenance (summary JSON):
dataset/population : 686 setups = 594 leads + 92 follow-ons; binary 571
(167 WIN / 404 LOSS)
feature_sha16 : 0414e401522ea4e2 (unchanged)
label_sha16 : b52a80677d5e9b99 (unchanged)
schema_sha16 : 470193a9d372c348 (unchanged)
model config : LogisticRegression(C=1.0, L2, max_iter=5000), seed 42
folds : ((0,300,300,395),(0,395,395,490),(0,490,490,571))
purge gaps : 555 / 472 / 321 bars
Git commit : recorded at run time
```
## M. Final feature classification
```text
Deterministic classification (thresholds documented in the script; NOT tuned
to results):
STABLE : direction, h4_gate, m30_gate (one dimension, all -),
choch_age_bars (all -), zone_width_atr (all -),
price_in_zone_offset, dist_to_zone_center_atr
(one affine dimension, all +).
VARIABLE : choch_latency_bars (all +, magnitude spread),
atr_at_entry (all +, F3 spike).
INCONCLUSIVE : sweep_age_bars, zone_age_bars (sign flips),
zone_type_code (constant, zero coefficient).
FOCUS features (top-4 by |std coef| in >= 2 folds, stable sign, mean|std| >=
0.05): direction/h4/m30 (CONTEXT) and price_in_zone_offset /
dist_to_zone_center_atr (ZONE). After collapsing exact duplicates these are
TWO distinct dimensions: (1) ZONE entry position, (2) CONTEXT direction/gate.
FINAL CLASSIFICATION : D — MIXED CONTRIBUTION
- based on coefficient sign consistency (stable in both leading
dimensions), magnitude consistency, fold stability (ranks 1-2 all folds),
semantic plausibility (zone-geometry + direction context), and feature
distribution stability (no fold-specific distribution driver for F2);
- NOT based on highest AUC, a single fold, or a single large coefficient.
Deserves further investigation? YES, cautiously: the ZONE entry-position
dimension and the CONTEXT direction dimension are the only candidates with
rank-consistent, sign-stable, fold-stable contribution to the weak signal.
The evidence remains a weak ranking edge (P3-S20: pooled OOS ROC 0.579), so
further investigation should be descriptive/mechanistic or an explicitly
authorized restrained-nonlinear study — NOT feature removal, NOT deployment.
```
## N. Production files unchanged
```text
NO MQL5 change. Unchanged: MQL5/** production files, F1/F2/F3/F4 runtime
semantics, Candidate Setup, FEATURE_CONTRACT.md, SniperGold_ML.mqh, model
weights, legacy MLP. Research-only artifacts under ml/p3/baseline/ only.
```
## O. Label contract unchanged
```text
P3_S16_LABEL_CONTRACT v1 : UNCHANGED (WIN=TP-before-SL, LOSS=SL-before-TP,
UNRESOLVED/AMBIGUOUS preserved; TP=+1.5 ATR, SL=-0.75 ATR, H=16). No label
v2 in this session. No label issue requiring documentation was found.
```
## P. ML escalation status
```text
NOT escalated. No calibration (that is P3-S21.2), no tree/boost/MLP/LSTM/
Informer/regime/ensemble, no feature ablation, no permutation importance, no
feature selection, no threshold/HP tuning, no retraining with subsets. Only
the EXISTING frozen logistic's coefficients were described.
```
## Q. Forge commits
```text
1. test: define P3-S21.1 feature contribution audit coverage (S21.1-T01..T07)
2. research: audit frozen Logistic feature contributions (P3-S21.1)
3. docs: record P3-S21.1 feature contribution results (this report)
P3_S21_1_FINAL_SHA = <recorded after push>
```
## R. Final Forge HEAD
```text
See Provenance (verified local == remote == P3_S21_1_FINAL_SHA after push).
```
## S. Working tree
```text
CLEAN after push (verified).
```
## T. P3-S21.2 readiness
```text
P3-S21.2 (CALIBRATION DIAGNOSTICS) is NOT started automatically. This
session's finding (D — MIXED CONTRIBUTION, ZONE-leading + CONTEXT secondary,
weak but rank/sign stable) is recorded for the owner decision. P3-S21.2
remains a SEPARATELY AUTHORIZED phase (probability calibration diagnostics,
NOT threshold tuning). Any feature removal / ablation / nonlinear model /
external data / production change also requires a separate authorization.
```
---
## Provenance / evidence
```text
Scripts (research-only, new in P3-S21.1):
ml/p3/baseline/p3_s21_1_feature_audit.py (frozen refit + verification +
inventory + coefficients +
groups + stats + redundancy)
ml/p3/baseline/test_p3_s21_1_feature_audit.py (S21.1-T01..T07; 7/7 PASS)
Outputs (ml/p3/baseline/output/):
p3_s21_1_feature_inventory.csv
p3_s21_1_feature_coefficients.csv
p3_s21_1_fold_coefficients.csv
p3_s21_1_group_contributions.csv
p3_s21_1_fold_feature_stats.csv
p3_s21_1_summary.json
p3_s21_1_tests.json
Reused (UNCHANGED):
ml/p3/baseline/prepare_dataset.py (12-feature schema, label v1)
ml/p3/baseline/walk_forward.py (frozen folds + config, imported only)
ml/p3/baseline/output/p3_s20_oos_predictions.csv (reproduction oracle)
ml/p3/baseline/output/p3_s20_summary.json (pooled ROC oracle)
Regression : S21.1-T01..T07 7/7 PASS; guard scan clean; no historical
artifact modified or weakened.
```
*End of P3-S21.1 feature contribution audit. Verdict: D — MIXED CONTRIBUTION
(ZONE-leading: entry-position + width; CONTEXT secondary: direction/gate;
both sign-stable across the three pre-registered folds; weak ranking edge,
honest caveats). Exact reproduction of the P3-S20 frozen logistic was
verified before any coefficient was reported.*
Copy permalink