Mean Reversion Basics

The opposite premise

Every strategy you've learned so far has been some flavor of trend-following: buy strength, sell weakness, ride the move. Mean reversion says the opposite: when price stretches far from its average, it tends to snap back. Buy weakness, sell strength. Fade extremes.

Kaufman sets up the dichotomy cleanly:

Trend-following and momentum techniques will buy and sell in the direction of the price move, mean-reverting strategies do just the opposite. — Perry Kaufman

These aren't competing worldviews — they're regime-specific. The same market can be trend-following at one timescale and mean-reverting at another; different markets have different natural regimes.

The trade profile — inverted

Kaufman's comparison table (kaufman.txt:41180) is the cleanest summary you'll find:

Same expected-value math, inverted distribution. A trend-follower tolerates being wrong 70% of the time because the 30% of winners pay for it all. A mean-reverter tolerates giving back occasional large losses because 80% of trades are small winners.

The critical asymmetry: catastrophic losers in mean reversion don't just hurt — they destroy. LTCM in 1998 was a mean-reversion fund that worked 99% of the time, until it didn't.

When mean reversion works

Kaufman names the conditions:

A market that has high noise is good for mean-reverting and arbitrage strategies. One with low noise favors trend-following.

And more specifically:

Equity indexes have a tendency to be mean reverting because they have a large amount of price noise.

Short timeframes favor mean reversion too:

Short-term traders focus on mean reversion or fast directional price moves.

Rule of thumb:

• Index ETFs on intraday timeframes → mean reversion often works
• Single stocks on multi-day timeframes → trend-following often works
• Commodities long-term → trend-following dominates
• FX pairs in calm regimes → mean reversion viable, until a macro shock

Common implementations

Bollinger band fade

Fade the 2σ extremes: buy at the lower band, sell at the upper. We covered Bollinger mechanics in its own lesson; the mean-reversion interpretation is one of two ways to use it (breakout-trend is the other).

RSI / Stochastic extreme fade

Buy when RSI < 30, sell when RSI > 70. Same trap as the Bollinger fade: works in ranges, dies in trends (oscillators can pin at 70+ or 30− for extended stretches in strong trends).

Z-score fade

Compute z = (price − mean(N)) / σ(N). Fade z > +2 (short), z < −2 (long). Target: z = 0 (return to mean).

Pair trading

Two historically-correlated instruments diverge. Long the one that's lagging, short the one that's led. Exit when the correlation re-converges. Cointegration is the rigorous mathematical framework (Engle-Granger 1987, Johansen 1991). Kaufman devotes much of chapter 14 to this.

ETF cross-sectional mean reversion

Rank instruments in a basket (S&P sectors, country ETFs) by recent performance. Short the winners, long the losers, expecting reversion. Kaufman documents this at kaufman.txt:24513.

Play with it

Flip the market to sideways and watch the fade strategy rack up winners. Then flip to up or down and watch it get murdered — in a trending market, price walks along the outer band and never returns to the mean before stopping out. The win-rate collapse is dramatic.

Bump the σ stretch from 2 to 3 and winners come faster (price is more extreme so the mean-reversion payoff is shorter) but trades are rarer. Drop to 1σ and trades fire constantly but the edge per trade drops.

The LTCM story — the textbook failure

Kaufman's single LTCM reference (kaufman.txt:39689) is also the best thing written on what goes wrong with mean reversion:

Long-Term Capital made the mistake of rationalizing a few large losses, convincing themselves that the combination of events that caused the largest equity drawdowns during historic testing would not occur again. If those losses were removed, then the expected risk was much lower, trading could be leveraged, and profits would be increased proportionally. They were right and wrong. The same events did not happen again, but different circumstances resulted in losses just as large. Because of their high leverage and massive positions, they shook the entire financial system when they collapsed. If anything, history does not show enough risk.

Read the last sentence three times: "history does not show enough risk." This is the core epistemic problem with mean-reversion backtests. They work until they meet an event the historical sample never saw — at which point the leverage that came from "this pattern has never failed before" becomes the detonator.

LTCM was run by two Nobel-laureate economists and the former head of Salomon's bond arb desk. They blew up in ~4 months in 1998 on correlations that had held for decades.

The Efficiency Ratio — Kaufman's regime detector

Kaufman provides a concrete tool for deciding when to mean-revert vs trend-follow. The Efficiency Ratio (ER):

Numerator: net price change over N bars. Denominator: sum of absolute bar-to-bar changes.

ER = 1 → pure trend (all price change was in one direction) ER = 0 → pure noise (net change is zero)

Typical window N = 8–10. Kaufman (kaufman.txt:43811): "the efficiency ratio can be a valuable tool for selecting the right market for a specific strategy."

Rule of thumb:

• ER > 0.5 → trending regime → favor trend-following
• ER < 0.3 → ranging / noisy regime → favor mean reversion
• Between → neither system has edge

Across 30 world futures markets 1998–2011, Kaufman found Eurodollars and Short Sterling had the highest ER (most trending) while FTSE, EuroStoxx, and Russell 2000 had the lowest (most mean-reverting). Market choice matters as much as strategy choice.

The academic short-term reversal anomaly

Separate from strategy-trading, there's a well-documented academic finding: short-horizon (1-week to 1-month) stock returns show negative serial correlation. Stocks that were up last week tend to underperform next week, controlling for market direction.

Originally documented by Jegadeesh (1990) and Lehmann (1990). Not in Kaufman directly — he cites neither — but the empirical pattern supports the short-timeframe mean-reversion story. Hedge funds have been trading this systematically since at least 2000 (typically with factor exposures stripped out).

Kaufman does document related behavior in index ETFs 1998–2018 (kaufman.txt:40137): non-random run lengths where "every market (except gold) had a run of 13 or more" — consistent with serial correlation exploitable by mean-reverting systems on short timeframes.

Hidden traps

• Leverage. The single most common way mean reversion blows up. Small winners tempt you to size up, because past draws don't seem to punish you. Then the tail event arrives at 5×–10× size.
• Tight stops. Mean reversion needs room. If your stop is at 1σ but your fade entry is at 2σ, you're selling into exactly the continuation you were hoping would reverse. Very-large or no-stop configurations are typical; Kaufman: "Only very large stop-loss can be used."
• Running in trending markets. Kaufman, again: "Profit-taking is essential… High volatility is an advantage… Low volatility should be avoided." The converse — running in a quiet trend — is death.
• Ignoring correlation during crisis. Pairs that have historically mean-reverted can stay divergent for weeks during a shock. LTCM assumed sovereign-bond spreads would converge; 1998 Russia default broke the assumption and their leverage killed them.
• Backtest survivor bias. Historical tests overweight pairs that eventually worked. Real-world, the pairs that stayed broken aren't in your dataset.
• Thinking ~80% win rate = good system. The win rate is the trap. What matters is win_rate × avg_win vs (1-win_rate) × avg_loss. Mean reversion's big-loss tail can easily negate the 80% win rate in a single event.

Quick check

What you now know

• Mean reversion is the opposite trade profile to trend-following: ~80% win rate, small winners, occasional catastrophic loser. Inverted distribution of trend-following's ~30% / large winner / small loser profile.
• Best conditions: high-noise markets (equity indexes), short timeframes, high volatility.
• Worst conditions: low-noise trending markets (commodities in steady regimes). Price walks the band; mean is never revisited.
• Common implementations: Bollinger fade, z-score fade, RSI extreme fade, pair trading (cointegration), ETF cross-sectional ranking.
• The Efficiency Ratio (ER) = |net change| / |total movement| over N bars tells you the regime. ER > 0.5 = trend-follow; ER < 0.3 = mean-revert.
• The LTCM lesson: "history does not show enough risk." Leverage turns occasional tail losses into extinction events.
• Leverage is the destroyer. High win rate tempts sizing up; the tail event then kills. Mean-reversion traders who survive keep sizes small.

Next up: System Testing Basics — what a backtest actually measures, and (more importantly) what it misses.