Mean Reversion and Momentum
Definition
A trading strategy can only be profitable if security prices are either mean-reverting (they return to a mean level) or trending / momentum (they keep moving in the same direction). These are the two organizing regimes of statistical arbitrage. Notes here follow Ernest Chan’s Quantitative Trading, chapters 7–8.
Core Ideas
The dichotomy
- Mean reversion — the spread of a pair or portfolio returns to its mean. Cross-sectional mean reversion (relative to a group of stocks) happens much more often than time-series mean reversion. Mean-reverting regimes are more prevalent than trending ones.
- Momentum — generated by the slow diffusion of information: as more people learn news or a large institutional order executes incrementally, price is driven the same way. A canonical example is Post-Earnings-Announcement Drift (PEAD).
Effect of competition differs by type:
- Mean reversion → arbitrage opportunities are gradually eliminated, returns decay to zero, and a rising share of signals turn out to be genuine fundamental repricing (won’t revert).
- Momentum → the time horizon over which the trend continues shrinks as news spreads faster; trades entered after equilibrium are unprofitable.
Stationarity and cointegration
- A time series is stationary (“integrated of order zero”) if it never drifts farther and farther from its initial value. A stationary price series guarantees a mean-reversion strategy is profitable as long as stationarity persists (not guaranteed). The converse is false — even non-stationary series have exploitable short-term reversals.
- Two non-stationary price series are cointegrated if a long/short combination of them is stationary. Test with the cointegrating augmented Dickey-Fuller (CADF) test.
- Cointegration ≠ correlation: correlation is about the short-term co-movement of returns; cointegration is about the long-term behaviour of prices. Two stocks can be cointegrated (their weighted spread always returns to a level) yet have uncorrelated daily returns.
Ornstein-Uhlenbeck and half-life
Mean reversion of a spread z(t) is modeled by the Ornstein-Uhlenbeck equation:
where μ is the long-run mean, dW is Gaussian noise, and θ is the rate of reversion. From θ you compute the half-life — a robust estimate of the optimal holding period for a reversion trade. (Momentum holding periods are error-prone by contrast, due to few signals.)
Regime change and Conditional Parameter Optimization (CPO)
Bull/bear markets are regimes. In CPO, the trader’s own strategy still makes the predictions; machine learning only optimizes the strategy’s parameters daily (e.g. lookback, entry threshold, weights) using technical-indicator features (Bollinger Z-score, Money Flow, Force Index, Donchian Channel, ATR, Awesome Oscillator, ADX). This keeps transparency and interpretability — unlike naïvely predicting next-day returns with ML.
Exit strategy
Entry signals are strategy-specific; exits use one of: a fixed holding period, a target price / profit cap, the latest entry signals, or a stop price. Stop losses suit momentum strategies but hurt reversal strategies (a reverting spread that moved against you is more likely to revert).
Seasonal strategies
Calendar effects like the January effect (worst-performing small-caps rebound in January) have become unprofitable in equities. Commodity-futures seasonal trades (gasoline, natural gas) remain profitable but occur only once a year — vulnerable to data-snooping bias and, for natural gas, extreme volatility (Amaranth’s $6B loss).
High-frequency trading
If a high Sharpe Ratio is the goal, trade at high frequency rather than holding overnight. By the law of large numbers, more bets → smaller percent deviation from mean return → higher Sharpe. HFT strategies generate the highest long-term compounded growth but are very hard to backtest and technology-reliant. Related: a highly-leveraged low-beta portfolio should out-compound an unleveraged high-beta one (low-beta stocks empirically carry higher Sharpe).
Relationships
- Quantitative Trading — the discipline this sits inside; backtest biases and risk management
- Options Strategies — options statistical arbitrage exploits the same cointegration / mean-reversion logic
- Factor Model (Fama-French) — factor models rely on factor-return momentum to work
- Sharpe Ratio — the law-of-large-numbers rationale for high-frequency trading
- Kelly Criterion — leverage on low-beta portfolios
- Trading & Finance — parent topic
See Also
- Dow Theory — where trend-following originates: a trend persists until a clear reversal signal