The Epps Effect

Correlation between two stocks should, in principle, be a stable property of how similarly their fundamentals move, whatever timescale you measure it over. Epps observed the opposite in real data: measure two stocks' correlation using 1-minute returns and it comes out noticeably lower than measuring the same pair's correlation using daily or monthly returns. As the measurement interval shrinks toward the scale of individual trades, the estimated correlation decays toward zero — a pattern now called the Epps effect.

The cause is not that stocks genuinely move independently minute to minute and then coordinate over longer horizons; it's an artifact of two microstructure realities. First, trades in different stocks don't happen at perfectly synchronized moments, so aligning "1-minute returns" for two names really compares slightly stale, asynchronous prices, which mechanically weakens measured correlation. Second, bid-ask bounce and other short-horizon noise add return variance that has nothing to do with common information, diluting any genuine shared signal that is present.

This matters directly for anyone estimating covariances from high-frequency data — a pairs-trading or statistical-arbitrage model built on 1-minute correlations will understate true co-movement unless the data is desynchronization-corrected or aggregated to a coarser interval first.