Mean reversion systems are trading strategies built on the premise that a price or, more often, a price spread tends to oscillate around a stable long-run average, so that extreme deviations from that average are temporary and likely to be corrected. The trader sells (or fades) what is unusually high and buys what is unusually low, betting on the snap-back rather than the continuation. The core tension is structural and unforgiving: the strategy is right most of the time and wrong rarely, but each rare error can be large — a series that "should" revert may instead be regime-shifting, trending, or breaking down permanently. Mean reversion therefore earns small, frequent gains in exchange for occasional, fat-tailed losses, the mirror image of trend following.

How it's calculated / formed

Mean reversion systems differ mainly in what is assumed to revert and how the deviation is measured.

• Single-series (overbought/oversold) reversion. A signal fires when a price stretches far from a moving average or an oscillator hits an extreme. Larry Connors' RSI(2) is the canonical example: in stocks above their 200-day SMA, buy when the 2-period RSI drops below ~5 (sometimes 10), exit when RSI(2) rises above ~65–75 (QuantifiedStrategies, StratBase). Bollinger Bands (price vs. a band of ±2 standard deviations) and a rolling z-score of price are the other common single-series triggers (QuantInsti).

• Spread / pairs reversion (statistical arbitrage). Rather than a single price, the system trades the spread between two (or more) related assets, which can be far more reliably stationary than any single stock. The spread is formed by cointegration — a regression giving a hedge ratio such that the residual series is stationary. Stationarity is verified with an Augmented Dickey-Fuller (ADF) or Johansen test; importantly, cointegration is a long-run relationship, whereas simple correlation captures only co-movement and can mislead (Hudson & Thames docs). Entry/exit are then z-score thresholds on the spread (e.g., enter at ±2σ, exit near 0).

• Continuous-time modeling. The spread is often modeled as an Ornstein-Uhlenbeck (OU) process, dX = θ(μ − X)dt + σ dW, where θ is the speed of reversion. From θ comes the half-life of reversion, ln(2)/θ — the expected time to close half the gap. Practitioners commonly treat half-lives under ~30 days as tradeable and those beyond ~90 days as too slow to recover round-trip costs (AI Fin Hub). The OU framework also supports optimal entry/exit and stop-loss boundaries that account for transaction costs (Leung & Li, arXiv 1411.5062).

The seminal "distance method" of Gatev, Goetzmann & Rouwenhorst is deliberately simpler: match pairs by minimum sum-of-squared-deviations between normalized prices over a 12-month formation window, then trade when prices diverge by more than 2 historical standard deviations over the following 6 months (GGR working paper, NBER w7032).

How it's used in practice

The academic literature documents a horizon pattern: prices tend to reverse at short horizons (the weekly/monthly short-term reversal of Jegadeesh 1990 and Lehmann 1990) and again at very long horizons (the multi-year overreaction reversal of De Bondt & Thaler 1985), while momentum tends to dominate the intermediate 3–12 month band (Jegadeesh & Titman 1993). This is why short-term, market-timing systems (RSI-2, gap fades, "buy the dip above the 200-day") lean mean-reverting, while medium-term trend systems lean the other way. Note that these are distinct, separately documented effects, not one universal law of any single price series.

Practical use is dominated by a few disciplines:

• Regime / trend filter. Connors' rules only buy oversold readings above the 200-day average — fading dips inside an uptrend, never standing in front of a downtrend.
• Market-neutral construction. Pairs and stat-arb books pair longs and shorts (or build factor-residual portfolios à la Avellaneda & Lee 2008) to strip out market beta, so the bet is purely on the spread.
• Position sizing and exit discipline are the whole game. Because losses are fat-tailed, size is kept small and exits are time- or level-based. Notably, Connors found that adding fixed stop-losses reduced backtested S&P 500 returns from the 1990s–2010, because stops fired right before the bounce (QuantifiedStrategies) — a genuine and uncomfortable tension between equity-curve smoothing and the strategy's own logic.

Adoption, debate & evidence

Mean reversion is one of the most heavily researched and widely deployed quant families, but the evidence is a story of decay. GGR documented average annualized excess returns up to ~11% (roughly 12% in the 1962–1997 sample) on top pairs, net of conservative cost estimates (Oxford RFS). However, Do & Faff (2010) showed a clear declining trend in U.S. pairs-trading profitability after the 1980s, accelerating in the 2000s — consistent with the strategy being arbitraged away as it became widely known and computing cheap.

The OU sophistication does not reliably beat the crude approach: at least one study found OU-based pairs trading underperformed the naive distance algorithm, a caution that theoretical elegance does not equal P&L (arXiv 2412.12458). For single-series signals, RSI-2 backtests report high win rates (sources cite >75% in some equity tests), but win rate is a treacherous metric for a strategy whose risk lives in the losing tail — and most such backtests are short, U.S.-equity, long-only-the-dip, and vulnerable to the very stop-removal that flatters returns. Treat "edge" claims here as period- and market-specific, not laws.

Strengths & limitations

It works best in range-bound, liquid, range-respecting markets and on economically linked spreads (same sector, dual-listings, ETF-vs-basket) where there is a real arbitrage force pulling the spread back. It fails — sometimes catastrophically — when the assumed equilibrium breaks: a merger, default, index reconstitution, or regime shift turns a "cheap" leg into a permanently broken one. The #1 misuse is fading a genuine trend or a structural break — repeatedly buying a falling knife because an oscillator looks oversold, while the half-life quietly goes infinite. Cointegration that held in-sample frequently fails out-of-sample, so the relationship must be re-tested continuously, and the strategy is unusually sensitive to transaction costs and short-borrow availability.

Sources

• Gatev, Goetzmann & Rouwenhorst, "Pairs Trading: Performance of a Relative-Value Arbitrage Rule" — NBER w7032, Oxford RFS 2006
• Do & Faff (2010), declining pairs-trading profitability — summarized in Warsaw WP 2/2022
• Leung & Li, "Optimal Mean Reversion Trading with Transaction Costs and Stop-Loss Exit" — arXiv 1411.5062
• OU vs. naive pairs trading underperformance — arXiv 2412.12458
• Connors RSI(2) rules, stop-loss effect, win rates — QuantifiedStrategies, StratBase
• Cointegration vs. correlation, ADF/Johansen, half-life — Hudson & Thames arbitragelab, QuantInsti, AI Fin Hub half-life
• Horizon/regime split (short & long = reversion, mid = momentum) — short-term reversal: Jegadeesh (1990, Journal of Finance) and Lehmann (1990, QJE); long-term overreaction reversal: De Bondt & Thaler (1985, Journal of Finance); intermediate momentum: Jegadeesh & Titman (1993, Journal of Finance)

Disputes flagged: magnitude and persistence of mean-reversion "edge" is contested — strong in-sample (GGR), decaying out-of-sample (Do & Faff); RSI-2 win-rate claims come from period-specific, mostly U.S.-equity backtests and should not be read as a stable edge.