Add a strict holdout/walk-forward + baseline gate to factor-weight learning,
per the P4 guardrail: learned weights are never auto-applied until minimum
sample, holdout/walk-forward, and baseline comparison all pass.
- backend/app/weight_learning.py:
- FactorLearning gained ic_train / ic_holdout / validated / gate_notes.
- apply_validation_gate(...) splits a chronological IC series into train +
holdout and only marks validated=True when: total >= MIN_SAMPLE_PERIODS,
each window >= its minimum, train AND holdout IC are positive (beat the
BASELINE_IC=0) and agree in sign, and the pooled |t| > MIN_IC_TSTAT.
- apply_weight_update now keeps new_weight == old_weight for any factor
that is not validated (no auto-apply); only validated factors move.
- learn_momentum_gated(...) builds PIT momentum ICs then applies the gate.
- backend/app/__init__.py: /api/v1/learning/momentum uses the gated learner
and surfaces ic_train/ic_holdout/validated/gate_notes.
- tests: gate (16) via rewritten suite — full backend 286 passed.
Live probe on current price archive: validated=false with
gate_note 'IC not above baseline (0.0711/-0.1143)' — momentum is not
validated, weight stays unchanged (new_weight=None).
A. Cross-theme comparability:
- compute_theme_surprises now weight-normalizes by total |weight| (weighted
average), so every theme surprise on same [-1,1] scale regardless of factor
count/weight (retail 0.189->0.145; auto_credit 1.0->0.64).
B. Historical factor store (enables learning macro/demographic factors):
- New factor_history.py: append-only per-factor JSONL, dedupes unchanged
values, rejects non-finite, records every FACTORS value each scheduler run.
- scheduler.py: jobs carry fetch_module; refresh_all records factor history
(non-fatal); added bank_npl job.
- GET /api/v1/learning/factors?min_points= reports n_points/learnable per
factor so users see when P4 learning unlocks (validated query parsing).
- weight_learning: generic learn_factor_series() aggregator (momentum reuses).
Independent review deleg_5dd358e3 passed=true (empty security/logic arrays);
its two robustness suggestions applied (finite guard in record(), clean 400 on
bad min_points). 234 tests pass; Vite build passes.
P0-B (registry is the single source of truth for scoring):
- FACTORS now carries center/span normalization spec; unused hand-written
per-theme surprise blocks in dashboard.py replaced by one registry-driven
compute_theme_surprises() (themes.py).
- THEMES['banks'] adds bank_npl weight so NPL is genuinely blended.
- factor_value/normalize hardened against NaN/inf (finite guards).
- Board re-ranks (TRUE/GULF up, TOP->3) per registry weights; 3 new tests
incl. 'changing a registry weight changes output'.
P3 (point-in-time backtest):
- run_backtest is now a real multi-rebalance engine (reallocates every window,
reconciles holdings, marks to market) instead of allocate-once+break.
- Added leakage_guard (False unless a PIT score_fn is supplied), planned vs
actual rebalances, and momentum_at() true 12-1 (skips last month, PIT).
P4 (factor-weight learning):
- weight_learning.py: cross-sectional Spearman IC, forward-return builder,
IC aggregation + t-stat, and apply_weight_update (new = clip(old*(1+shrink*IC))).
- GET /api/v1/learning/momentum endpoint. Live result: momentum IC=0.012
t=0.132 over 22 periods -> momentum has no reliable predictive power here.
Macro/demographic factors blocked (no historical factor vintages yet).
Two independent review gates passed (deleg_fe6f45cd, deleg_718218f8): empty
security/logic arrays; their non-blocking suggestions applied (finite guards,
dedupe leakage_guard resolution). 226 tests pass; Vite build passes.
