Skip to content
All library documents

Epps Effect Bias in High-Frequency Return Correlations

Article Quant Q&A · Author: vanguard2k
Read the original

Summary

The document discusses the Epps effect: measured correlations between assets can fall as return sampling becomes more frequent, in part because trades occur asynchronously and market microstructure affects observations. The answer says the effect persists and cautions that there is no single time horizon at which assets suddenly decouple. It cites a study reporting correlation estimates at one-minute and five-minute sampling intervals that were respectively about 60% and 90% of the underlying value.

It describes a proposed correction based on differences between correlations, which can cancel the bias when tick-level data are available. With only broker-supplied minute bars, the respondent expects the effect to be harder to remove. The answer tentatively favors the additional resolution of minute data while recommending that its correlations account for the bias. The cited observations use older data, and the document explicitly notes that market conditions may have changed, so the figures should not be treated as current universal thresholds.

Key ideas

  • Measured cross-asset correlations may decline at finer sampling intervals because of asynchronous trading and microstructure effects.
  • The effect is gradual rather than tied to a universal decoupling time.
  • The cited study found stronger correlation bias at one-minute intervals than at five-minute intervals.
  • A correlation-difference method may reduce the bias when tick data are available.
  • The cited data are dated, so the reported magnitudes may not reflect current markets.

Tags

StatisticsMarket microstructureEquities
Full text
# When does the Epps effect start?


# When does the Epps effect start?












Wikipedia defines the Epps effect as follows:

In econometrics and time series analysis, the Epps effect, named after T. W. Epps, is the phenomenon that the empirical correlation between the returns of two different stocks decreases as the sampling frequency of data increases. The phenomenon is caused by non-synchronous/asynchronous trading and discretization effects.

Epps wrote this paper in 1979 and the speed of information propagation has increased dramatically since.

Does this effect still exist? If it does, what is the time horizon in practice when does the "decoupling" of assets happen now?

## Answer by H. Arponen (score 7, accepted)

https://quant.stackexchange.com/a/9183

I think it's alive and well. I don't think there's a specific "decoupling" time, but if you look at e.g. Munnix et al. "Statistical causes for the Epps effect in microstructure noise", it seems that the biased correlation is about 60% of the real value for 1 min data and about 90% for 5 min data, so you could say that 5 min is pretty safe, but 1 min is probably not.

FYI, for example Zhang et al. "Estimating Covariation: Epps Effect, Microstructure Noise" present a fairly straightforward way to effectively remove the Epps effect by considering a certain difference of correlations such that the Epps effect/ bias cancels. It works if you happen to have the real tick data at hand, but if you only have, say, 1 min data from a broker etc. then I guess you pretty much have to live with the Epps effect.

I think it is still better to use 1 min data instead of e.g. 5 min data due to the increased resolution (though I'm not sure). It's just that for 1 min data the correlations could be better estimated by taking the Epps effect into account.

EDIT: By the way, the data in the Munnix paper seemed to be from 2007, so maybe things have changed since then...

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.