feat(factors): P3-5 candidate factor pack, validated through the P3-4 matrix

[StackOverFlow11]

Summary

P3-5 — factor candidate pack (EXPLORATORY). Adds a conservative, daily, PIT-safe candidate factor set and re-runs the unchanged P3-4 robustness matrix to test whether the legacy trio's weak signal (P3-3/P3-4 finding) was just a too-narrow factor set. No alpha / portfolio / execution / OOS / robustness model change. Not tuned. Not a return claim.

Candidates (all unit-locked: math vs manual reference, leading-window NaN, future-bar perturbation invariance, per-symbol isolation, immutability)

Factor	Definition	PIT argument
reversal_5/20	−(close[t]/close[t−w] − 1)	exact negative of momentum (reuses its machinery; ≤t bars)
volatility_20	trailing 20d return std (ddof=1, full window only)	≤t bars
liquidity_20	log(mean(amount, 20)); non-positive → NaN (never −inf)	same-day bar amount
value_ep / value_bp	1/pe, 1/pb from one daily_basic fetch; ≤0 → NaN	ratios published same-day
grossprofit_margin	quality field → joins SUPPORTED_FIELDS	existing ann_date as-of machinery, unchanged

Wiring: dispatch extended (reversal* / volatility* / liquidity* / value fields); a window-name/params mismatch is a readable config error (a spec named reversal_5 with window: 10 can never silently mislabel a column); _maybe_enrich_value wired into phase0 / phase2 / OOS-cell paths; demo + value/financial factors still fail readably. Legacy configs reproduce their numbers (demo ic 0.96 / annual 0.84, test-locked).

Config: config/phase3_real_factor_candidates.yaml = legacy trio + 7 candidates through the same matrix cells as P3-4 (CSI300×2020-2022 stays an explicit, disclosed skip_cells entry). Old-vs-new reads off ONE run: raw-factor ICs are per-column, so the trio's per-cell ICs double as a no-drift cross-check.

Real matrix (3 cells, 2.3h; secret scan clean)

No-drift smoke ✓: the legacy trio's per-cell ICs are identical to the P3-4 report (momentum 0.0017/0.0062/0.0151; roe −0.0181/0.0066/0.0015; np_yoy −0.0117/0.0046/0.0070).

Cross-cell stability (test IC > 0 / sign consistent):

Series	test IC > 0	sign consistent	test IC range
value_ep	3/3	2/3	+0.037 … +0.052
value_bp	3/3	2/3	+0.044 … +0.056
volatility_20	0/3 (i.e. 3/3 negative)	2/3	−0.044 … −0.079
liquidity_20	0/3 (3/3 negative)	2/3	−0.011 … −0.036
overnight_mom_20	2/3	2/3	−0.010 … +0.016
reversal_5/20, grossprofit_margin	sign-flip	≤1/3	no signal
legacy trio	(P3-4 unchanged)	≤1/3	|IC| ≤ 0.015
combo_ic_weighted	3/3	3/3	+0.001 … +0.040

(table = the 11-factor rerun; all 10 prior series' per-cell ICs are unchanged vs the 10-factor run — per-column independence re-confirmed.)
| combo_equal_weight | 0/3 | 1/3 | −0.013 … −0.023 |

Findings (honest read):

1. The too-narrow hypothesis is SUPPORTED: conservative value (E/P, B/P) and low-volatility carry cross-cell-stable direction at an order of magnitude above the legacy trio.
2. The walk-forward ic_weighted combo picks the new signal up (3/3 positive + consistent); the all-factor equal-weight combo is diluted by the sign-flipping factors (0/3; CSI300 test annual −25.2% in the 11-factor run) — equal weight is sensitive to factor-set quality.
3. Portfolio (11-factor run): ic test −2.21% / −5.02% / −2.76%, beats equal weight in 3/3 cells (eq: −4.30% / −6.83% / −25.16%); CSI300 train +23.11% → test −2.76% — train→test decay remains.
4. Caveats disclosed in the report and docs: value/low-vol were a known strong A-share regime in 2020-2024; the three cells overlap in time (not independent samples); no cost-sensitivity analysis yet. Next step: an independent re-check of a value+lowvol subset.

Quality gates

Gate	Result
pytest -p no:cacheprovider	317 passed (was 300: +17 candidates; 1 legacy test reworked for the new mismatch guard)
ruff check .	clean
validate-config ×8 (all old + factor_candidates)	OK
run-phase0 demo	ic 0.96 / annual 0.84 unchanged
P3-3/P3-4 no-drift	trio per-cell ICs identical to the P3-4 report
Real matrix	SSE50×2 folds + CSI300×2022-2024 run; CSI300×2020-2022 explicit skip, disclosed
Secret scan	report + tree contain no token / .config.json content

Tests (network-free)

test_candidate_factors.py (17): per-factor math vs manual references, leading-window NaN, future-bar invariance, per-symbol isolation, non-positive guards, missing-column readable errors, ValueFactor contracts, grossprofit_margin in SUPPORTED_FIELDS, dispatch for every candidate + mismatch guard, value enrichment (one fetch, 1/pe / 1/pb math, non-positive → NaN, demo error, immutability), demo e2e with price candidates, candidates config validation.

Docs

RUNBOOK (P3-5 section incl. the one-run old-vs-new design and drop_missing note), TEST_REPORT (317 + breakdown), CLAUDE.md / AGENTS.md (verbatim copy) progress incl. findings + caveats.

Addendum: overnight_mom_20 (added on review request)

overnight_ret[t] = log(open[t]/close[t-1]), overnight_mom_20 = trailing 20-day sum. PIT: open[t] is known at the t open, factor computed at the t close; one front-adjust anchor keeps the ratio ex-dividend-safe; the t−1 close never crosses symbols; non-positive prices → NaN. +5 unit tests (manual-reference math, future-bar invariance, per-symbol isolation, guards) → 322 passed. Real 11-factor matrix rerun (2h; one CSI300 network failure → full clean rerun): test IC 2/3 positive (+0.008 / +0.016; negative in 2020-2022) — moderate-weak; all prior series' ICs unchanged.