Guide

Mean reversion trading explained

Harbor Capital's sector-rotation desk noticed a pattern: when the technology-sector ETF traded more than two standard deviations below its 60-day moving average relative to the broad market, it tended to outperform the S&P 500 over the next five to ten sessions — not because tech was “cheap,” but because short-term panic selling often overshot fundamentals and snapped back. A systematic mean reversion sleeve that bought those dislocations and exited at the moving average added 1.8% annualized alpha net of costs over three years, with a Sharpe near 1.1. Mean reversion trading assumes prices, spreads, or ratios that deviate from a historical baseline will return toward that baseline — buying “too low” and selling “too high” relative to a defined mean. It is the philosophical opposite of momentum, which bets that winners keep winning. Both can work; the edge lives in knowing which regime you are in and sizing for the speed of reversion. This guide explains the statistics behind mean reversion, common signal constructions, regime filters, a Harbor Capital sector-ETF worked example, a strategy decision table, pitfalls, and a production checklist.

What mean reversion means in markets

In finance, mean reversion is the tendency of a variable — a stock price, a spread between two ETFs, an RSI reading, a volatility index — to fluctuate around a central level and return after extreme moves. The “mean” might be a simple moving average, a fair-value model, a cointegrating relationship between two assets, or the long-run average of a ratio like price-to-earnings.

Mean reversion is not a law of physics. Structural breaks (bankruptcy, paradigm shifts, permanent de-rating) can turn a temporary dip into a permanent loss. Successful mean reversion traders therefore combine entry signals (how far from the mean?) with regime filters (is reversion still plausible?) and risk controls (what if the mean itself moves?).

Mean reversion vs momentum

Momentum: past winners continue outperforming over weeks to months. Works in trending macro environments and for factor premia like cross-sectional equity momentum.
Mean reversion: extreme short-term moves reverse. Works in range-bound markets, liquid large-caps, and relative-value spreads between linked securities.
Time horizon matters: daily returns show weak mean reversion at very short horizons and momentum at 3–12 month horizons in many equity markets. Your signal window must match the horizon where reversion actually occurs.

See technical analysis fundamentals for how trend-following and oscillator tools map to these two philosophies.

Statistical foundations

Rigorous mean reversion work starts with stationarity: does the series tend to return to a stable distribution, or is it a random walk that drifts forever? The Augmented Dickey-Fuller (ADF) test and Hurst exponent are common checks. A Hurst exponent below 0.5 suggests anti-persistent (mean-reverting) behavior; above 0.5 suggests trending.

Z-scores and bands

The workhorse signal is the z-score: how many standard deviations is the current value from its rolling mean?

z = (x_t - μ) / σ

Entry rules often look like: buy when z < -2, sell or cover when z crosses zero or +1. Bollinger bands are a visual version: price touching the lower band is roughly a −2σ event under Gaussian assumptions (which returns rarely satisfy — fat tails mean −2σ happens more often than a bell curve predicts).

Half-life and Ornstein-Uhlenbeck

For spreads and pairs, practitioners fit an Ornstein-Uhlenbeck (OU) process to estimate half-life — how many days until a deviation decays halfway back to the mean. If half-life is 12 days but you hold for 60, you are fighting slow reversion and paying carry. Pairs traders use this heavily; see pairs trading explained for spread construction and cointegration.

Common signal families

Price vs moving average: deviation from 20/50/200-day MA; works on indices and liquid ETFs more than single small-caps.
Oscillators: RSI < 30 (oversold), Stochastic extremes — mean reversion on the indicator, not always on price.
Cross-sectional: buy recent losers, sell recent winners within a universe (short-term reversal factor).
Spread / ratio: sector ETF vs benchmark, gold vs miners, near-month vs far-month futures — relative mean reversion.

Building a mean reversion strategy

A production-ready sleeve has five layers:

Universe. Liquid names with tight spreads — mean reversion dies in illiquid gaps. Prefer ETFs, large-caps, or cointegrated pairs.
Signal. Define the mean, lookback, and entry z-score threshold. Document whether you trade absolute price or a spread.
Regime filter. Skip entries when VIX is spiking, when ADX shows strong trend, or when macro events break historical relationships. Filters reduce trade count but cut tail losses.
Exit. Target the mean (z = 0), a partial reversion (z = −0.5), time stop (exit after N days regardless), or stop-loss if z widens further (the trade is wrong).
Sizing. Scale inversely with estimated variance or use Kelly-style fractional sizing from backtested win rate and payoff ratio — never full Kelly on noisy short-horizon signals.

Every rule should be validated with walk-forward backtesting that includes realistic costs, slippage, and out-of-sample periods. Mean reversion backtests overfit easily: many parameter combos look great in-sample on the same decade of range-bound data.

Worked example: Harbor Capital sector-ETF sleeve

Harbor Capital's quant team tested a relative mean reversion sleeve on eleven S&P sector ETFs (XLK, XLF, XLE, etc.) against SPY from 2018–2025.

Signal definition

Compute each sector ETF's excess return vs SPY over the prior 60 trading days.
Standardize using a rolling 252-day window of that excess return series.
Entry: when standardized excess < −1.75, buy the sector ETF and hedge with a beta-neutral SPY short sized via 60-day rolling beta.
Exit: when standardized excess crosses −0.25, or after 15 trading days, or stop if excess < −3.0 (regime break).

Results and lessons

Out-of-sample (2022–2025), the sleeve generated 142 trades with a 58% win rate and average winner 1.4x average loser. Best months clustered around macro scares (banking stress, rate-shock selloffs) when sector rotation overshot. Worst months coincided with persistent trends — energy outperforming for quarters during a supply shock, where “oversold” kept getting more oversold. The team added a simple filter: skip new entries when the sector's 50-day MA slope exceeded ±2% per month, cutting drawdown by 22% at the cost of 15% fewer trades.

Key takeaway: the edge was in relative dislocations within a diversified benchmark, not in calling absolute market bottoms.

Strategy decision table

Your situation	Favored approach	Caution
Liquid ETF vs benchmark dislocation	Relative z-score with beta hedge	Macro regime shifts can extend dislocations for months
Two cointegrated stocks	Pairs trade on spread z-score	Test cointegration rolling; relationships break on M&A
Intraday index futures	Short-horizon OU on order-flow imbalances	Requires low latency and tight cost model
Single small-cap stock “oversold”	Usually avoid — weak stationarity	High risk of value trap or delisting
Crypto alt vs BTC ratio	Ratio mean reversion with strict stops	Fat tails, 24/7 gaps, liquidity evaporation in stress
Strong trending market (ADX > 30)	Reduce or pause mean reversion	Fighting momentum is the classic blow-up mode

Pitfalls

Catching falling knives. A −3σ move often becomes −5σ when fundamentals break. Always pair z-score entries with stop rules.
Ignoring transaction costs. Short-horizon mean reversion with thin edges dies after commissions, spread, and market impact. Model costs in backtests.
Non-stationary means. A 200-day MA in a bubble is not a “fair value” anchor. Consider detrending or using relative spreads.
Overfitting lookback windows. The optimal 17-day RSI that worked 2015–2019 may fail 2020–2025. Walk-forward validation is mandatory.
Capacity limits. Cross-sectional reversal in micro-caps does not scale to billions. Know where your edge saturates.
Confusing oscillator reversion with price reversion. RSI can normalize while price keeps falling in a grind lower.
Leverage on slow half-life trades. Borrowing to hold a spread that mean-reverts over 90 days can erase gains with funding costs.

Production checklist

Define the mean, lookback, and whether you trade absolute or relative series.
Test stationarity (ADF, Hurst) on in-sample and rolling out-of-sample windows.
Specify entry z-score, exit target, time stop, and stop-loss widening threshold.
Add at least one regime filter (trend strength, volatility, macro calendar).
Backtest with realistic slippage, fees, and borrow costs for shorts.
Run walk-forward or purged cross-validation to reduce data snooping.
Estimate half-life for spread strategies; align hold period with reversion speed.
Size positions with risk management rules — cap single-trade loss at a fixed % of NAV.
Monitor live hit rate vs backtest; decay often signals regime change.
Document when to shut the sleeve off (drawdown limit, correlation spike to benchmark).

Key takeaways

Mean reversion bets on prices or spreads returning toward a baseline — the opposite of momentum, but both coexist across time horizons.
Z-scores, Bollinger bands, and OU half-life are the standard toolkit; stationarity tests tell you whether the bet is structurally plausible.
Regime filters separate viable entries from value traps — trending markets punish naive mean reversion.
Relative dislocations (sector vs index, pairs spreads) often behave more predictably than absolute single-stock bottoms.
Costs, capacity, and overfitting kill more mean reversion strategies than bad math.