Serial Correlation
Serial correlation (also called autocorrelation) is the statistical measure of the relationship between a variable's value at one point in time and its value at a previous point in time, quantified by the autocorrelation coefficient ranging from -1 (perfect negative serial correlation) to +1 (perfect positive serial correlation). In financial contexts, serial correlation in asset returns has profound implications for momentum and mean-reversion strategies, risk measurement, and the validity of performance metrics.
Key takeaways
- Positive serial correlation in returns means past positive returns predict future positive returns—the foundation of momentum strategies.
- Negative serial correlation means past positive returns predict future negative returns—mean-reversion strategies exploit this tendency in equity markets at short (daily) horizons.
- Hedge fund returns often exhibit positive serial correlation due to illiquid, hard-to-price assets (private equity, real estate, structured credit), causing artificially smooth NAV series and understated risk metrics.
- The Ljung-Box and Box-Pierce tests are standard statistical tests for the presence of serial correlation in return series.
- Autocorrelation-adjusted Sharpe ratios are higher than standard Sharpe ratios for positively autocorrelated return series—reflecting the understated volatility in smoothed return series.
Explanation
Serial correlation in financial time series is a departure from the random walk hypothesis—the idealized assumption that asset price changes are independent and identically distributed (i.i.d.). While the efficient market hypothesis in its weak form implies that past prices contain no information about future prices (zero serial correlation in returns), empirical evidence documents systematic serial correlation at multiple horizons: positive autocorrelation at 3–12 months (momentum effect), negative autocorrelation at very short (intraday, daily) horizons for liquid markets (bid-ask bounce), and positive then negative patterns at longer horizons (mean reversion after momentum).
The measurement of serial correlation in return series employs the autocorrelation function (ACF): ρ(k) = Cov(r_t, r_{t-k}) / Var(r_t), where k is the lag order. A first-order autocorrelation of 0.15 means that 15% of one period's return can be predicted from the prior period's return. While this seems modest in economic terms, it can be practically significant for trading strategies operating at scale. The Ljung-Box Q-statistic tests the joint null hypothesis that a set of autocorrelations (up to lag k) are all zero; rejection of this null indicates statistically significant serial dependence in the series.
Hedge fund serial correlation has been extensively studied as a red flag for performance reporting integrity. Getmansky, Lo, and Makarov (2004) documented that many hedge funds exhibit surprisingly high first-order autocorrelations in monthly returns—sometimes 0.3–0.4—in strategies that theoretically should generate near-zero autocorrelation (market-neutral, arbitrage, trend-following). They attributed this to 'return smoothing': either the deliberate marking of illiquid positions at stale prices (carrying them at cost or last quoted price rather than fair market value) or the gradual recognition of trading profits across multiple periods. The consequence is that the fund's reported Sharpe ratio significantly overstates true risk-adjusted performance—the volatility denominator is artificially compressed by the serial correlation.
Time series momentum—one of the most documented factor premia in academic finance—directly exploits positive serial correlation. Moskowitz, Ooi, and Pedersen (2012) showed that across 58 liquid futures markets (equities, currencies, commodities, fixed income), past 12-month returns positively predict next 12-month returns with statistical significance, producing a Sharpe ratio of approximately 1.3 from a long-short time series momentum strategy. The economic rationale combines under-reaction to information (markets adjust slowly to fundamental shifts, creating trending dynamics) and feedback effects (stop-loss selling and momentum strategies amplify price moves, sustaining trends beyond initial information).
Serial correlation also affects risk measurement. Standard volatility and VaR calculations assume i.i.d. returns; when returns are positively autocorrelated, multi-period volatility scales faster than the √T rule (it scales by a factor that exceeds √T when autocorrelation is positive), meaning that monthly VaR cannot simply be computed by multiplying daily VaR by √21. The correct scaling requires incorporating the autocorrelation structure into the variance-covariance estimation, typically through HAC (heteroskedasticity and autocorrelation consistent) estimators such as Newey-West. For portfolios with significant serial correlation in returns (e.g., private equity, CTA trend-followers), standard risk metrics must be corrected for serial dependence to accurately reflect true risk.
Formula
Autocorrelation: ρ(k) = Cov(r_t, r_{t-k}) / Var(r_t); Autocorrelation-Adjusted Sharpe: SR_adj = SR × √((1-ρ)/(1+ρ))
Example
A hedge fund manager reports the following 12 monthly returns: +1.5%, +2.0%, +1.8%, +1.2%, +1.9%, +2.1%, +1.4%, +1.7%, +2.2%, +1.6%, +1.8%, +1.9%. The series is suspiciously smooth—low volatility and consistent direction. Computing the first-order autocorrelation: the correlation between the 11 pairs of adjacent monthly returns is approximately 0.42—well above the 0.18 threshold for statistical significance at the 95% level in a 12-observation sample. An institutional investor investigating this fund would question whether the smooth returns reflect genuine portfolio performance or illiquid, hard-to-value positions being marked at stale prices. Applying the Getmansky-Lo-Makarov correction to unsmooth the returns produces an implied true monthly return standard deviation of 1.5% (rather than the reported 0.3%), reducing the apparent Sharpe ratio from 4.5 to 0.9—a significant deterioration that would affect the allocation decision. The investor requests independent portfolio-level valuation verification before committing capital.
Related terms
Alpha Signal Arbitrage Autocorrelation Correlation Covariance Efficient Market Hypothesis Equity Gradient Boosting Hedge Fund Itos Lemma Mean Reversion Private Equity