Feature Audit Report — SoccerPredictAI¶

Date: 2026-04-28 Auditor: GitHub Copilot (Claude Opus 4.7) — /skill-ml-system-audit full (audit 02/12) Scope: Feature pipeline — engineering, leakage, contract, train ↔ inference consistency Baseline: docs/validation/20260424/02_feature_audit.md

Delta vs baseline¶

No source files under src/features/, src/pipelines/features.py, src/pipelines/inference.py, or params.yaml modified since 2026-04-26. Baseline findings remain in force.

Confirmed inventory¶

4 rolling groups × 5 metrics × 5 windows × 3 sides + ELO ⇒ 363 columns in features.parquet.
Model input via select_model_features(): side="diff", window_sizes=[1,3], include_elo=True → diff_* (w1, w3) + diff_elo_pre + sex.
Single contract: features_meta.parquet (built from rolling cols via parse_feature(), ELO appended afterwards — order verified safe).

Train ↔ inference consistency¶

Path	Code reuse
Training (`features.py`)	`build_team_match_table` + `add_rolling_features` + `to_match_level` + `compute_elo_ratings` + `select_model_features`
Batch inference (`inference.py`)	Same functions — full parity ✅
Online `POST /predict/` (`tasks/predict.py`)	No server-side recomputation — client must supply pre-computed features

Leakage analysis (re-confirmed)¶

Rolling: .shift(1) before .rolling().sum() → no current-match leakage. Sort key [teamId, startTimeUtc, id] deterministic.
ELO: pre-match ratings only, documented in elo.py docstring.
Target: outcome_1x2 in build_team_match_table only via shifted history.

✅ No leakage detected.

Risk register (re-confirmed)¶

ID	Severity	Description	Status
F-01	P1	Online `POST /predict/` relies on client to supply features — no server-side feature contract enforcement	Open
F-02	P1	No runtime check that `classification.window_sizes ⊆ features.window_sizes` — silent NaN risk	Open
F-03	P2	`select_model_features()` API exposes `include_h2h` / `include_rest_days` but these features are not implemented	Open
F-04	P2	`coverage_w{n}` columns produced but never selected into model input	Open
F-05	P3	No input schema validation in `compute_all_match_features()` — missing columns become silent NaN/0	Open

Summary¶

Aspect	Status
Rolling features (4 groups) — train/inference parity	✅
ELO — train/inference parity, no leakage	✅
Single feature contract (`features_meta.parquet`)	✅
H2H / rest_days advertised in API	❌ not implemented
Server-side feature recomputation for online predict	❌
Runtime guard on window_sizes subset	❌

Recommendation: address F-01 (server-side feature lookup or strict request schema validation) and F-02 (assertion at pipeline init) — both protect production from quiet failures.

See baseline §1–§5 for code-level detail.