Mean Reversion Trading Deep Dive: From Galton 1886 to Pairs Trading

I want to write about mean reversion the way I wish someone had written about it for me. Not the textbook version, where it’s introduced as a single idea and then dropped. The actual lineage. Where it came from, why it works when it works, and the very specific moments where it stops working and traders lose money in surprising ways.

This is going to be a longer read than most blog posts. The reason is that mean reversion is one of the foundational ideas in quantitative trading, and most online explanations of it are bad. They either reduce it to “buy low, sell high” (which isn’t what mean reversion means) or they go straight to indicators without explaining what the indicators are measuring. I’d rather give you the full picture.

For Korean market readers specifically, this matters because Korean equities exhibit some of the cleaner mean-reversion behavior in developed markets, partly because of structural retail flow and partly because the KOSPI 200 universe is small enough that statistical relationships hold up well. I’ll come back to that toward the end.

A quick disclaimer: I’m not a licensed financial advisor and nothing here is investment advice. I’m writing this because the topic is fun and the existing English-language coverage is uneven.

The trade nobody really teaches you

Most retail investors hear “buy low, sell high” and assume that’s mean reversion. It isn’t, quite. The actual idea is statistical: when a price moves far enough away from its statistical average, the probability of it moving back toward that average is higher than the probability of it continuing further away. The keyword is “statistical.” This is not a story about value or about fundamentals. It’s a story about distributions.

Where this gets useful is when you can quantify “far enough away.” That’s where the indicators come in (RSI, stochastics, Bollinger Bands, z-scores), and that’s where the academic literature lives. The whole field of mean reversion trading is, in a sense, a long argument about how to measure distance from the mean and how to act on it.

The frustrating thing about mean reversion is that it works for a while, then stops working, then starts working again. Markets cycle between mean-reverting regimes and trending regimes, and a strategy that’s printing money in one regime quietly bleeds in the other. Understanding which regime you’re in is the harder problem than picking the right indicator.

So what is mean reversion, actually?

Formally, a process is mean reverting if its expected value at any future time is closer to its long-run mean than its current value is. In practice, for asset prices, you don’t observe the “true” mean. You estimate it, usually from a moving average. The choice of how to estimate that mean turns out to be most of the game.

The classic statistical version comes from Francis Galton’s 1886 paper on “regression toward mediocrity,” which was about human heights but accidentally founded the whole concept that gives us linear regression. Galton observed that tall parents tended to have children who were also tall but less so, and short parents had children who were short but less so. Extreme observations regressed toward the population mean over generations.

Markets do something similar but for different reasons. When a stock moves dramatically in one direction over a short period, two things tend to happen. Profit-takers sell into the move, and contrarian buyers (or short-sellers) step in expecting reversal. The combined effect is a damping force on extreme moves. It doesn’t always work, but it works often enough to be statistically detectable.

This statistical detectability is the entire reason there’s an industry around it.

The academic timeline that built the case

The financial literature on mean reversion really starts in the 1980s, when researchers began testing whether stock returns showed predictable reversal patterns. The key papers are worth knowing because they shape how every modern mean reversion strategy is framed.

De Bondt and Thaler (1985) published “Does the Stock Market Overreact?” in the Journal of Finance, and it became one of the most cited papers in behavioral finance. They built portfolios of past three-year “losers” and past three-year “winners” and found that the losers outperformed the winners over the subsequent three years by about 25 percentage points on average. The market overreacted to extreme news, and that overreaction got corrected over multi-year horizons. This is long-term mean reversion.

Lo and MacKinlay (1988) in “Stock Market Prices Do Not Follow Random Walks” applied variance ratio tests to weekly and daily US stock returns from 1962 to 1985 and rejected the random walk hypothesis. They found short-horizon negative autocorrelation, which is the statistical signature of mean reversion. This was the empirical foundation for thinking that mean reversion was real and not just a story.

Jegadeesh (1990) in “Evidence of Predictable Behavior of Security Returns” documented monthly reversal patterns. He sorted stocks into deciles by their previous month’s returns and found that the top decile underperformed the bottom decile by about 2.49% per month over the following month. That’s an enormous signal, although transaction costs and bid-ask spreads eat into it significantly.

Lehmann (1990) in the Quarterly Journal of Economics found similar weekly reversal effects with the title “Fads, Martingales, and Market Efficiency.” Same general phenomenon, different time horizon.

These four papers, plus the broader behavioral finance literature, established that mean reversion exists in equity markets at multiple time horizons, that it’s stronger for individual stocks than for indices, and that it varies with market conditions.

The four indicators traders actually use

The academic case is one thing. The practical question is how to translate “stocks mean-revert” into actual trades. Four indicators dominate.

RSI (Relative Strength Index) was introduced by J. Welles Wilder Jr. in his 1978 book “New Concepts in Technical Trading Systems.” The default 14-period RSI ranges from 0 to 100, with 70 traditionally interpreted as overbought and 30 as oversold. The math compares average gains to average losses over a window. The classic mean reversion trade is to short when RSI exceeds 70 and buy when it drops below 30. Wilder’s original work has held up surprisingly well.

Bollinger Bands, developed by John Bollinger in the 1980s, plot a 20-period moving average plus and minus two standard deviations. A move outside the bands is, by the math of the normal distribution, a roughly 5% probability event. Traders fade those moves expecting reversion to the middle band. The trick is that the standard deviation widens during trending periods, so the bands themselves move with the price.

Z-scores are the most explicitly statistical of the four. You calculate a rolling mean and rolling standard deviation, then express the current price as Z = (price minus mean) divided by standard deviation. A z-score of plus or minus 2 corresponds to the same 95% probability band as Bollinger. The advantage of explicit z-scores is that you can apply them to anything (price spreads between two stocks, ratios, transformed series), which makes them the workhorse of quantitative implementation.

Stochastic Oscillator, developed by George Lane in the late 1950s, measures where the current close sits relative to the high-low range over a lookback window. Lane found that closing prices tend to cluster near the highs in uptrends and near the lows in downtrends, and that this clustering signals impending reversals. Stochastics typically signal overbought above 80 and oversold below 20.

The honest truth is that all four indicators are measuring roughly the same thing from different angles. If you use them all together, you mostly get confirmation, not independent signal. Pick one, understand it deeply, and don’t pretend the others are adding much.

Pairs trading — the elegant version

The most elegant application of mean reversion is pairs trading. The idea is straightforward. Take two stocks whose prices have historically moved together (think highly correlated within the same industry). Construct a spread, usually as the price ratio or the price difference adjusted for some hedge ratio. When the spread deviates significantly from its mean, you go long the underperformer and short the outperformer, expecting the spread to converge.

The strategy was pioneered in the early 1980s at Morgan Stanley by a team led by Nunzio Tartaglia. Gerry Bamberger, who was on the team, is often credited with the original implementation. The Morgan Stanley quant group used early computers to monitor thousands of stock pairs in real time, executing automatically when spreads exceeded statistical thresholds. The group reportedly earned exceptional returns through the mid-1980s before being broken up.

The beauty of pairs trading is that it’s market-neutral by construction. If the overall market crashes, both stocks in the pair fall together, and the spread stays roughly constant. You only lose if the spread itself widens further (because of company-specific news, for example). This is why pairs trading historically had high Sharpe ratios even during equity bear markets.

In Korean markets, classic pairs include Samsung Electronics common against Samsung Electronics preferred, Hyundai Motor common against Hyundai Motor preferred, and various same-industry pairs like KB Financial and Shinhan. The preferred-common spread trade in particular has been a textbook mean reversion setup for decades, with dividends accruing to short-side and arbitrage discounts narrowing on corporate events.

The structural reason Korean preferred shares mean-revert against their common counterparts is that they pay the same dividends but have no voting rights, so the discount is essentially the value the market assigns to voting rights, and that value fluctuates within a fairly narrow band over time.

Why mean reversion stops working sometimes

This is the part nobody wants to talk about. Mean reversion strategies work until they don’t, and the failures are usually expensive.

The simplest failure mode is regime change. A stock that’s been mean-reverting for years starts trending strongly because of a fundamental change (new product, regulatory shift, M&A activity), and mean reversion bets keep buying the dips that don’t stop dipping. Sears, GE during certain stretches, and most cyclical industrials in commodity downturns are examples where mean reversion traders got run over.

The second failure mode is what quants call “fat tails.” Real financial return distributions have far more extreme outcomes than a normal distribution would predict. A 4-sigma move under a normal distribution should happen roughly once every 15,000 trading days. In actual markets, 4-sigma moves happen multiple times per decade. So z-score-based strategies that assume normality systematically underestimate tail risk.

The third failure mode is correlation breakdown in pairs trading. Two stocks that have moved together for years can suddenly diverge for company-specific reasons (one gets acquired, one has an accounting scandal, one’s industry shifts). When that happens, the pairs spread keeps widening rather than reverting, and the losses can be unbounded.

The fourth failure mode, and the most subtle, is academic decay. Papers like Jegadeesh (1990) reported huge returns to monthly reversal strategies. But returns to these strategies in the post-2000 era have been much smaller, partly because quantitative funds have arbitraged away most of the effect, and partly because microstructure has changed (decimalization, electronic trading, ETFs). A strategy that’s well known is a strategy that probably doesn’t work as well as the back-tests suggest.

The combined message is that mean reversion is real but unstable, profitable on average but lethal in tails, and increasingly arbitraged in modern markets. None of which means it doesn’t work. It means it requires more humility than the textbook treatment suggests.

What I’d actually do with this

I’ll close with what I tell myself when I’m thinking about applying mean reversion in practice.

First, regime matters more than indicator. Spend more time deciding whether the market or the specific stock is in a mean-reverting or trending regime than on tuning the indicator. Simple regime tests (rolling autocorrelation, rolling variance ratio) outperform fancy indicators in most studies I’ve seen.

Second, position sizing matters more than entry signal. The expected return on a mean reversion trade is usually small relative to the worst-case loss. So sizing each trade so that a bad outcome doesn’t end your year is more important than perfecting the entry.

Third, the pairs trading family is generally better than single-name mean reversion. The market-neutral construction reduces the regime-change risk, although it introduces correlation-breakdown risk. For most traders without industrial-strength infrastructure, fading extreme moves in individual stocks is usually a losing game over time, while careful pairs trading can produce stable returns.

Fourth, Korean markets specifically reward mean reversion in the preferred-common pair space. If you have access to KRX with the ability to short, the structural discount-narrowing trade has been one of the more reliable setups in Korean equities over the past two decades. It requires patience, since dividend dates and corporate events drive a lot of the convergence.

And finally, accept that mean reversion is a slow, statistical game. The trades don’t pay 50% in a month. They pay small fractional returns repeatedly, and the discipline is in not abandoning the strategy during the periods when it underperforms. Most traders fail at mean reversion not because the math is wrong but because they can’t sit through the drawdowns.

References

• De Bondt, W. F. M., and Thaler, R. H. (1985). Does the stock market overreact? Journal of Finance, 40(3), 793–805.
• Lo, A. W., and MacKinlay, A. C. (1988). Stock market prices do not follow random walks: Evidence from a simple specification test. Review of Financial Studies, 1(1), 41–66.
• Jegadeesh, N. (1990). Evidence of predictable behavior of security returns. Journal of Finance, 45(3), 881–898.
• Lehmann, B. N. (1990). Fads, martingales, and market efficiency. Quarterly Journal of Economics, 105(1), 1–28.
• Wilder, J. W. Jr. (1978). New Concepts in Technical Trading Systems. Trend Research.
• Gatev, E., Goetzmann, W. N., and Rouwenhorst, K. G. (2006). Pairs trading: Performance of a relative-value arbitrage rule. Review of Financial Studies, 19(3), 797–827.

Disclaimer: This article is for educational purposes only and does not constitute investment advice. The author is not a licensed financial advisor. Mean reversion strategies have real risks including regime change, fat-tail losses, and academic-effect decay. Past statistical relationships do not guarantee future results. Consult a qualified financial advisor before making investment decisions.