Factor And Timeseries Research

Factor research and time-series forecasting

The libraries here are mostly healthy. The danger is that their default cross-validation and forward-return conventions do not purge anything, so a leaky result looks like a clean API call.

1. Factor evaluation

1.1 alphalens-reloaded — and its forward-return convention

alphalens-reloaded 0.4.6 (2025-06-02, Apache-2.0, 642★) is alive but low-velocity. The original quantopian/alphalens is dead (0.4.0, 2020) — do not use it.

🚨 Verified in source: compute_forward_returns does pct_change(period).shift(-period), so the forward return for date t starts at t's own price. It never lags your factor. If your factor is computed from date t's close, alphalens is scoring you as if you traded that same close.

Fix: lag the factor yourself before passing it in — factor.groupby(level=1).shift(1) — or build forward returns from the next open. The IC alphalens reports on an unlagged close-based factor is not achievable.

1.2 Qlib Alpha158 / Alpha360

pyqlib 0.9.7 (MIT, 48,255★). ✅ The label is Ref($close,-2)/Ref($close,-1)-1 — deliberately leakage-safe: it trades at T+1's close and measures to T+2, so the signal at T is never scored against a price it could see.

🚨 The real trap is normalization: ZScoreNorm is fit over fit_start_time..fit_end_time. Pass the full sample and you leak the test distribution into every feature, silently. Set the fit window to your training period only.

1.3 Alpha101 — a licensing problem, not a quality problem

🚨 The canonical Python port (yli188, 864★) and nearly every other Alpha101 repo has no licence file = all rights reserved. You cannot legally use or redistribute them. ✅ Menooker/KunQuant (Apache-2.0, active 2026-05) is the one safely usable implementation.

1.4 Symbolic alpha mining

gplearn was revived — 0.4.3 shipped 2026-01-07 after a 3.7-year gap; now needs Python ≥3.11 and sklearn ≥1.8. Notes calling it abandoned are stale.

🚨 Every expression the search evaluates is a trial. A genetic program that examines 10⁵ candidates has a trial count of 10⁵ for Deflated Sharpe purposes. See backtest-validation.

1.5 Fama-French and asset-pricing regressions

✅ Live-tested: Ken French's ZIPs return HTTP 200; a downloaded file parsed to 1,200 monthly rows through 202606. pandas-datareader 0.11.1 (2026-06-24) is itself revived and its famafrench reader works (its parser was silently wrong before 0.11.0 — re-pull anything you fetched with 0.10.0). 🚨 getFamaFrenchFactors is dead (2 releases, both 2019-05-18).

linearmodels 7.0 (2025-10-21, NCSA licence, 1,066★) is healthy and remains the only real Fama-MacBeth / asset-pricing option. pyfixest 0.60.0 is fast and modern but has no Fama-MacBeth — complementary, not a replacement.

🚨 Fama-MacBeth with overlapping returns understates standard errors. Use Newey-West / HAC. With panel data, choose the clustering dimension deliberately — clustering on the wrong dimension is how a t-stat of 2 becomes a t-stat of 5.

1.6 Two negative results worth knowing

• 🚨 No credible open-source Barra-style risk model exists. Every replication found is unlicensed and abandoned (2017–2023). cvxgrp/cvxrisk (MIT, active) is credible but is a risk interface, not an estimated model — you still supply the factor exposures.
• 🚨 Event studies have no usable library. The only PyPI package, eventstudy, is 0.1a12 — an alpha from 2021, GPL-3.0, 69★; alternatives top out at 12★. Write it yourself. references/_event-study-method.md has the full methodology: market-model estimation window, event window, CAR vs BHAR, cross-sectional t-statistics, and Boehmer-Musumeci-Poulsen for event-induced variance (which is the correction almost everyone omits).

2. Forecasting

🚨 Prophet is maintained (1.4.0, 2026-08-15) but was among M5's worst performers and is actively harmful for financial series. Listed for recognition; not recommended.

2.1 🚨 Verified leakage in the forecasting CV splitters

• sktime's ExpandingWindowSplitter and SlidingWindowSplitter have NO gap / purge parameter. With a multi-step horizon, train and test overlap in the information they encode.
• darts.historical_forecasts has no purge either.
• 🔴 Worse: darts' retrain=False makes every backtest point in-sample — a one-word change turns an honest walk-forward into a fitted-on-everything curve.

For financial labels with any horizon, use purgedcv or skfolio's CombinatorialPurgedCV instead of the forecasting libraries' own splitters. See backtest-validation §2.

2.2 tsfresh's multiple-testing defaults are correct — don't "fix" them

✅ HYPOTHESES_INDEPENDENT = False by default, so select_features uses Benjamini-Yekutieli (fdr_by), which is valid under arbitrary dependence — not plain Benjamini-Hochberg. Financial features are heavily dependent, so this is the right choice. Flipping it to get more features through is exactly the mistake the parameter exists to prevent.

Separately: extracting features across the train/test boundary leaks. Extract per-fold.

2.3 Time-series foundation models — honest evidence

🚨 Licences — verified directly from the HuggingFace API, 2026-09-04. Do not assume Apache:

🚨 pip install tirex installs an unrelated 2022 dimensionality-reduction package — the real one is tirex-ts.

Does any of it work on prices?

• On GIFT-Eval, only Chronos-2 and TimesFM-2.5 beat classical AutoTheta at high frequency, and benchmark contamination is documented.
• The directly relevant finance study (arXiv 2607.05291: 9 TSFMs vs 8 econometric specifications, 50 assets, realized volatility) found only Tiny Time Mixers beat Log-HAR, narrowly — and Mincer-Zarnowitz recalibration showed most of that edge was better scaling, not better dynamics.

Conclusion: treat TSFMs as a baseline to beat, not a source of alpha. A HAR model on realized volatility remains hard to beat, and any TSFM claim should be checked against it.

3. The recurring traps

1. Unlagged factors in alphalens (§1.1) — the most common factor-research bug.
2. Splitters with no purge in sktime/darts (§2.1).
3. Normalizers fit on the full sample — Qlib's ZScoreNorm, any StandardScaler outside a Pipeline.
4. Overlapping-return standard errors in Fama-MacBeth — needs Newey-West.
5. Survivorship in the factor universe — a factor tested on today's index members is not a factor. See research-integrity-guards §1.
6. .shift() sign errors and resample label placement (label=, closed=).
7. Search-space size as trial count — gplearn, Optuna, AutoML. See backtest-validation.

4. Reference files

references/<library>.md for exact versions, licences, signatures and quirks; references/_event-study-method.md for the methodology you have to implement yourself.

Per-library deep dives

The optional fin-libraries plugin carries a dedicated skill for each library below. Load one only after this skill has told you which library you want:

• lib-alphalens — alphalens