Codebase analysis: findings report (no code changes)

[Copilot]

Deep audit of the ensemble branch (feature/comprehensive-improvements) to identify remaining issues after three prior audit rounds. No code was changed — this PR documents findings only.

Confirmed fixed (from prior audits)

All original P0 execution bugs resolved: per-position _update_equity, stale-position closing in reconcile_target_weights, equity snapshot before reconciliation loop, sell weight_change sign, purged walk-forward CV replacing iloc split, LightGBM early stopping, risk engine initialization, bracket leg preservation, bridge created once outside loop, fill verification polling.

Remaining issues found

High

• _cancel_existing_sl_tp_orders cancellation logic is inverted (live_bot.py:357): None is included in the order_class allowlist, so any sell-side order without an explicit class (regular limit sells, manual orders) gets cancelled. Only orders with a recognized bracket class (oco, oto, bracket) should be cancelled.

Medium

• Active feature columns not winsorized: mom_12_1, mr_zscore_60, sector_rel_mom are in FEATURE_COLS but absent from _WINSORIZE_COLS in features.py. Outliers in these flow raw into training and inference.
• HMM normal state runs at 70% exposure (hmm_regime.py:70): Permanent 30% cash drag in baseline market conditions. Makes A/B comparison with Bot A (full exposure) structurally unfair and depresses returns mechanically.
• Ensemble IC gate fails the whole ensemble when only stacking fails (ensemble.py:365): When _fit_meta_learner has insufficient OOS data, validation_ic stays 0.0 and the quality gate (< 0.02) triggers equal-weight fallback — even though the base models may have positive individual ICs.
• _has_traded_today fails closed silently on broker error (live_bot.py:522): Transient API errors at startup cause the bot to skip the cycle entirely with only a log line — no CRITICAL alert fired.
• ATR comment/code mismatch (live_bot.py:119–125): Comment documents 3x SL / 5x TP; live defaults are 2x / 3x.
• Confidence sizing uses min-max normalization (predict.py:768): Single outlier score collapses all others toward zero, concentrating conviction weight and undermining HRP diversification. Rank-based normalization is more robust.

Low

• No Bonferroni correction in drift detection: 11 KS tests at p=0.01 → ~10.5% false-positive rate per cycle.
• Module-level globals read by Telegram thread without locks (predict.py:83–88): _last_drift_report, _last_raw_ohlcv, _last_prediction_scores written by main loop, read by polling thread — no synchronization.
• HMM re-fetches SPY 3mo on every rebalance cycle (live_bot.py:1464): Redundant HTTP call; could slice the already-fetched 2y SPY data from initialization.
• MLflow run nesting risk (predict.py:1027): Drift start_run inside get_ml_weights could become a child of an unclosed training run on exception paths.
• base.py reconcile_portfolio target_weight=0 skip (original Finding #6): execution.py was fixed; BaseBroker.reconcile_portfolio was not reviewed — status unconfirmed.
• sector_rel_mom universe mismatch: Cross-sectional feature computed on a slightly different ticker universe at inference vs. training due to min-history filtering — structural (non-stochastic) drift.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
dashboard	Error		Mar 1, 2026 1:25am