Momentum

A momentum strategy buys assets that have outperformed recently and (often) sells the underperformers. Recently usually means the last 3 to 12 months. The canonical academic origin is Jegadeesh & Titman (1993), which showed that US equities sorted by past 3–12 month returns kept outperforming for the next 3–12 months. Every cross-sectional momentum implementation since is a variant of that idea.

for each asset in universe:
  score = price_today / price_N_months_ago - 1
rank universe by score
hold top K, rebalance monthly

The choice of N matters. A 3-month lookback is “recent momentum” — fast, high turnover, picks up the current leadership of the tape. The academic standard is 12-1: a 12-month lookback, but skip the most recent month. The skip exists because of short-term reversal — assets that ran hard in the last 4 weeks tend to mean-revert before continuing the longer trend. If you don't skip the recent month you'll buy local tops and get knifed.

Momentum earns money when the cross-section has strong, persistent leadership and the market regime stays put long enough for the rotation to compound. Concrete examples: tech megacaps in 2020–2021, energy in 2003–2007, crypto majors in 2017. The common thread is a multi-quarter regime where the same handful of names lead the tape and laggards keep lagging.

It also works best when the universe has low cross-correlation. Pick 5 names that are 90% correlated and your “winners” portfolio is just one bet. Pick 50 names across sectors and you're actually harvesting the cross-sectional spread.

Momentum gets crushed at regime changes. The mechanism is mechanical: yesterday's winners are by definition the most-bought, most-loved positions, so they have the most to give back when sentiment flips. February 2020, October 2008, March 2009, and the late-2022 rotation out of mega-cap tech are textbook momentum-crash windows.

This is also why the long-short version of momentum is famously fragile: the short leg (recent losers) is exactly where the rebound bid lands first. Long-only momentum survives crashes; long-short can blow up in a single week.

• Walk-forward windows. Momentum strategies often have one OOS window that crushes — the one that captured a persistent regime. If only one of three OOS windows is good, the headline OOS Sharpe is regime-specific, not validated edge. Look for consistency across windows, not for one heroic number.
• IC decay. If the momentum signal decays fast (decay ratio < 0.7 from horizon t to t+1), the factor is crowded out — by the time you trade on it, the alpha is gone. Slow-decaying ICs are the ones worth trading.
• Turnover. Cross-sectional momentum with monthly rebalances typically runs 100–300%/yr turnover. If your backtest shows 800%, you're churning and transaction costs will eat the edge in production.

• Testing only on the post-2009 bull run. Momentum looks magical from 2009–2021 because every dip was bought. Backtest through 2008 and 2020 too, or you're grading on a curve.
• Ignoring transaction costs. Momentum is a high-turnover style. A 1.5 gross Sharpe at 200%/yr turnover on a small universe drops to 0.7 net after costs in most realistic execution models.
• Forgetting the skip. The 12-month lookback without the 1-month skip is a different (worse) strategy. If you're writing the formula yourself, double-check this.

• Sharpe vs PSR vs DSR → the metrics that tell you whether your momentum Sharpe is real or selection inflation.
• Walk-forward vs CPCV → the two ways we estimate out-of-sample momentum performance.