The Epps Effect: Short-Interval Correlation and Asynchronous Prices
Why do high-frequency asset correlations fall as sampling intervals shrink?
Estimate cross-asset correlation at several sampling intervals while checking asynchronous trading, microstructure noise and estimator choice. A decline at short intervals can have multiple causes; synchronization methods may change the estimate without supplying a universal fix.
The measured pattern depends on instruments, observation times and horizon; an asynchronous estimator cannot remove every noise or market-hours bias.
Evidence map
| Aspect | Finding |
|---|---|
| What it is | In several high-frequency equity samples, measured correlation between assets falls as the return interval becomes shorter. |
| Key result / formula | Epps documented lower short-interval correlations in the stock data he studied. |
| Why it matters for backtesting | If a portfolio's historical correlations are estimated from short bars with stale or non-overlapping prices, the estimated comovement may be distorted. |
What it is
Asynchronous trades are one important mechanism: a price can stay unchanged in one instrument while another updates. The magnitude and mechanism depend on the instruments, trading intensity, sampling scheme and horizon. The observation does not imply that every short-interval correlation vanishes or that one estimator universally fixes it.
Key result / formula
When each observed price updates at a different time, a common calendar grid can pair one asset's new return with another asset's stale price. Hayashi and Yoshida propose a covariance estimator based on overlapping observation intervals rather than forcing a common grid. Its consistency comes with a specified asynchronous diffusion setting and sufficient observations; it does not remove every microstructure bias, noise source or market-hours mismatch. A researcher should distinguish trading-time sampling, stale quotes, bid-ask bounce and true changes in dependence before attributing a measured pattern to a single cause.
Why it matters for backtesting
That can affect a hedge ratio or diversification estimate, but the direction and magnitude must be measured for the actual data. Compare estimates over several bar widths on the same overlapping trading session; report each asset's fraction of unchanged or missing bars. A plateau can be informative, not proof that the longest interval is unbiased. If event-level timestamps are available, an asynchronous estimator can be compared with the bar-based result under its assumptions. If only bars are available, disclose that observation timing inside each bar remains unresolved. Do not automatically equate a low short-interval estimate with a profitable diversification opportunity.
Worked check
Imagine one asset trades every minute while another trades every fifteen minutes. A one-minute grid records many zero changes for the second asset, even while the first changes. Recompute correlation at one, five and fifteen minutes, holding session hours fixed, and count stale bars. If the estimates differ, investigate asynchronous observation before changing portfolio weights. A strategy-specific decision still needs out-of-sample risk and transaction-cost checks.
Source
Epps, "Comovements in Stock Prices in the Very Short Run", Journal of the American Statistical Association 74(366), 1979, 291-298; Hayashi & Yoshida, "On covariance estimation of non-synchronously observed diffusion processes", Bernoulli 11(2), 2005, 359-379. Primary source