Stock cross-sectional momentum — exploration
Does cross-sectional momentum — buying recent winners, shorting recent losers — carry a tradeable signal among individual US stocks? On a survivorship-free universe (delisted names included), the signal is real and monotonic, and lives in the market-neutral long-short, not in beating the market long-only. Sweet spot: a ~9-month lookback (skipping the last month) held ~3 months.
Signal strength — most significant configurations
| universe | lookback | hold | decile spread /yr | t | IC | long-short net /yr |
|---|---|---|---|---|---|---|
| top-3000 | 3-0 | 3mo | 6.3% | 3.3 | 0.013 | +3.1% |
| top-3000 | 3-1 | 3mo | 5.6% | 2.9 | 0.012 | +2.4% |
| top-3000 | 3-0 | 6mo | 4.3% | 2.8 | 0.016 | +2.7% |
| top-1000 | 3-0 | 3mo | 6.4% | 2.5 | 0.013 | +3.2% |
| top-500 | 9-0 | 3mo | 7.7% | 2.3 | 0.022 | +4.5% |
| top-3000 | 6-0 | 3mo | 5.4% | 2.2 | 0.016 | +2.2% |
Taming the momentum crash — variant comparison
| construction | net /yr | max DD | Sharpe |
|---|---|---|---|
| raw long-short | +2.3% | 68% | 0.22 |
| vol-target 3mo · cap1.5 | +7.6% | 45% | 0.43 |
| short-guard (½ short in stress) ★ | +8.4% | 33% | 0.53 |
| beta-neutral (matched pairs) | +2.2% | 65% | 0.21 |
| beta-neutral + vol-target 3mo | +4.6% | 64% | 0.31 |
Survivorship-free US common-stock history extends back to 2004, so this in-sample window includes the 2008–09 momentum crash. The crash-guard scales the long-short inversely to its own trailing volatility; the best-Sharpe refinement shown is short-guard (½ short in stress). Holdout 2023+ remains sealed. The long-only top decile lags SPY after cost, so a tradeable version is market-neutral (long winners / short losers). The raw long-short is uninvestable (~68% drawdown in the 2009 crash), so we tame it: faster volatility-targeting (scale down when the strategy's own vol spikes) already cuts the drawdown to ~45%, but the biggest win is a short-leg guard — halving the short position in high-volatility months, because the crash is driven by beaten-down losers rocketing in the rebound. That lifts the Sharpe from 0.22 to 0.53 and cuts the worst drawdown from 68% to 33%. In-sample, beta-neutralising the long-short (the "matched-pairs" idea) looked to barely matter — the average net beta was only −0.10. The sealed holdout overturned that: out-of-sample the naive long-short carried market beta ≈+1.0 (2023-24 momentum ran into high-beta mega-cap winners), so the unhedged version is not market-neutral when it counts. The frozen beta-neutral confirmation (D-006) cut realised holdout beta to +0.14 while keeping +17.7%/yr (Sharpe 1.04) — it cleared the gate. Net read: cross-sectional momentum is real and confirmed out-of-sample; its beta-neutral packaging clears the holdout too, though the in-sample neutral edge is thin, so this is a forward-paper candidate, not a live edge. See the frozen confirmations D-005 / D-006 in the discovery log. Winsorised returns, $5 price floor, net of 40 bps round-trip.