Serial Correlation Returns: When Past Returns Predict Future Returns

Key Takeaways

• Daily equity returns have lag-1 autocorrelation statistically indistinguishable from zero, while squared returns show ACF above 0.25 at lag 1.
• Lo and MacKinlay rejected the random-walk null for weekly US equity indices in 1988, finding modest but significant positive serial correlation.
• Illiquid assets like hedge fund and small-cap indices show artificially high positive autocorrelation from stale prices, not real momentum.
• Serial correlation guides strategy horizon: intraday reversal strategies exploit microstructure noise that disappears at daily and longer horizons.

What It Is

Serial correlation at lag k is the correlation of a return series with itself shifted by k periods:

rho(k) = Corr( r_t , r_{t-k} )

A value of zero at all lags is consistent with a random walk in prices, which is the baseline for weak-form market efficiency. Positive serial correlation indicates trend or momentum at that horizon. Negative serial correlation indicates mean reversion.

Fama's 1965 study of Dow stocks found daily serial correlations clustered tightly around zero but slightly positive. Lo and MacKinlay (1988) used a variance-ratio test to reject the random-walk null for weekly US equity indices, finding modest positive autocorrelation. Cont's 2001 survey treats "absence of linear autocorrelation in returns at short horizons" as a stylized fact that coexists with strong autocorrelation in squared returns.

The Intuition

If returns were strongly autocorrelated, a simple rule (buy after up days, sell after down days) would print money. Arbitrage pressure pushes daily autocorrelation toward zero in liquid markets. It does not push it all the way to zero because transaction costs, microstructure frictions, and risk premia leave small residual patterns.

Different horizons tell different stories:

• Intraday: Mean-reversion dominates at the one-minute to one-hour scale, driven by bid-ask bounce and short-term liquidity rebalancing.
• Daily: Near zero for liquid stocks and indices.
• Weekly to monthly: Mild positive autocorrelation (momentum) for individual stocks over 3 to 12 month windows.
• Multi-year: Mean reversion at 3 to 5 year horizons, documented by Fama and French and many others.

Meanwhile the autocorrelation of absolute or squared returns is strong and persistent across all these horizons. This is the volatility clustering mentioned in a separate article.

How It Works

The sample autocorrelation estimator is:

rho_hat(k) = sum_{t=k+1}^{T} (r_t - mean)(r_{t-k} - mean)
           / sum_{t=1}^{T} (r_t - mean)^2

For a white-noise series of length T, the standard error of each rho_hat(k) is approximately 1/sqrt(T). With 2,500 daily observations, anything between plus and minus 0.04 is statistically indistinguishable from zero at the 5 percent level.

The Ljung-Box Q-statistic aggregates autocorrelations across multiple lags into a single test:

Q = T(T+2) * sum_{k=1}^{h} rho_hat(k)^2 / (T - k)

Under the null of no autocorrelation, Q follows a chi-squared distribution with h degrees of freedom. A significant Q rejects the random-walk null. Practitioners typically run Ljung-Box on both raw returns and squared returns, expecting to fail to reject for the first and reject strongly for the second.

Worked Example

Take 10 years of daily SPY returns, roughly 2,520 observations. A typical sample produces:

rho(1) on returns      ~ -0.05
rho(5) on returns      ~ +0.02
Ljung-Box Q(20)        ~ 25   (p-value ~0.20, fail to reject)

rho(1) on |returns|    ~ +0.25
rho(20) on |returns|   ~ +0.15
Ljung-Box Q(20) on |r| ~ 1800 (p-value essentially zero)

The raw-return series looks indistinguishable from white noise at daily frequency. The volatility series is highly persistent. That split is the signature of modern equity market microstructure. It says prices are difficult to predict, but risk is very predictable.

Common Mistakes

• Claiming to find alpha from small daily autocorrelations. A rho(1) of minus 0.05 is not a strategy. Bid-ask spreads, commissions, and market impact on any realistic trade size consume signals of that size many times over.

• Running a single lag-1 test and stopping. Patterns can show up at lag 5, 20, or 250 even when lag 1 is silent. Ljung-Box across multiple lags is the standard diagnostic.

• Ignoring stale-price bias in illiquid assets. Hedge fund, real estate, and small-cap indices often show artificially high positive autocorrelation because their prices are smoothed or stale. Correcting for this (unsmoothing) is standard in due diligence.

• Confusing autocorrelation in returns with autocorrelation in volatility. The first is about predictable direction. The second is about predictable magnitude. Both can be tested with the same tools, but they mean very different things.

• Forgetting that efficient does not mean zero. Weak-form efficiency allows small, non-exploitable autocorrelations consistent with transaction costs. A statistically significant rho does not automatically mean an inefficient market; it may simply mean costs are priced in.

Frequently Asked Questions

Q: What is serial correlation in returns in simple terms? It measures whether knowing today's return tells you anything about tomorrow's return. In liquid equity markets at daily frequency the answer is almost nothing, but meaningful patterns exist at intraday, monthly, and multi-year horizons.

Q: How does serial correlation in returns affect investment decisions? It underpins the academic justification for both momentum strategies at monthly horizons, where mildly positive serial correlation provides the edge, and mean-reversion strategies at intraday horizons where microstructure creates predictable reversals.

Q: What is a real-world example of serial correlation in returns? The autocorrelation of absolute SPY daily returns at lag 1 is typically around 0.25 and stays positive for 20 or more lags, while the autocorrelation of the raw signed returns at those same lags is statistically indistinguishable from zero.

Q: How can investors avoid mistakes with serial correlation in returns? Run Ljung-Box tests across multiple lags rather than only lag 1, check for stale-price bias in illiquid asset returns before drawing conclusions about momentum, and remember that statistically significant autocorrelation is not automatically economically exploitable after trading costs.

Q: How is serial correlation in returns different from volatility clustering? Serial correlation in returns concerns whether the direction of price changes is predictable. Volatility clustering concerns whether the magnitude of price changes is predictable. Real markets show almost no direction predictability at daily frequency but strong magnitude predictability.

Sources

1. Fama, E. (1965). "The Behavior of Stock-Market Prices." Journal of Business, 38(1), 34-105. https://www.jstor.org/stable/2350752
2. Cont, R. (2001). "Empirical properties of asset returns: stylized facts and statistical issues." Quantitative Finance, 1(2), 223-236. https://www.tandfonline.com/doi/abs/10.1080/713665670
3. 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. https://doi.org/10.1093/rfs/1.1.41
4. Ljung, G.M. and Box, G.E.P. (1978). "On a Measure of Lack of Fit in Time Series Models." Biometrika, 65(2), 297-303. https://doi.org/10.1093/biomet/65.2.297

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.