Industry-relative momentum (at 4-week lookback)

Thesis (rejected at this horizon): Names that are the strongest in their sector over the trailing 4 weeks continue to outperform for the next 2-3 weeks. Asness 1994 / Moskowitz / Grinblatt 1999.

Code: research/industry_momentum_validate.py Report: research/industry_momentum_report.json (n=1.16M rows)

Why it failed: The 4-week lookback is too short. Within-sector ranking at this horizon picks up Jegadeesh 1990 short-term reversal, not momentum. The bottom quintile of relative momentum outperformed the top quintile across every horizon tested:

The opposite-sign trade (long bottom, short top) does have a small positive expectancy but the per-trade spread is too small (≤ 1 %) to be worth productizing after frictions.

Verdict: At a 4-week lookback the textbook signal is mean-reversion, not momentum.

Re-tested at 6-month and 12-month lookbacks (the canonical Jegadeesh setup)

Same direction at every lookback. Within-sector ranking is consistently mean-reverting, not momentum-following.

Cross-sector momentum does work at the 6-month horizon (see research/cross_sector_momentum_report.json: +2.2 pp at 126 d, t=1.7). So the canonical Jegadeesh momentum is largely a sector effect; when you neutralise for sector, what's left is short-term reversal. The practical takeaway: rank ETFs, not stocks within ETFs, when you want momentum exposure.