Autocorrelation of tick returns at different horizons
Posted: Tue Nov 03, 2026 5:40 am
Autocorrelation of tick returns at different horizons
At the tick level, returns tend to show negative autocorrelation due to bid-ask bounce. At a few seconds, the sign can flip as momentum from order flow persists. At a few minutes, it depends on the regime. The horizon at which a strategy trades determines whether it fights or follows the structure.
Measuring the autocorrelation of mid-price returns at different sampling intervals, say 1 second, 5 seconds, 30 seconds, and 5 minutes, shows where mean reversion and momentum dominate. The results vary by pair, time of day, and market conditions.
For a scalper, this informs the choice between fade and follow strategies. If 10-second returns show positive autocorrelation during the London open, momentum entries make sense. If they show negative autocorrelation in the Asian session, fading is preferable.
The numbers are small, often between 0.02 and 0.1, but small correlations can be exploited with enough trades if costs are low.
Compute them across weeks of data, and check stability. If the sign changes randomly, there is nothing to trade.
What do the autocorrelations look like at the horizon the strategy uses?
At the tick level, returns tend to show negative autocorrelation due to bid-ask bounce. At a few seconds, the sign can flip as momentum from order flow persists. At a few minutes, it depends on the regime. The horizon at which a strategy trades determines whether it fights or follows the structure.
Measuring the autocorrelation of mid-price returns at different sampling intervals, say 1 second, 5 seconds, 30 seconds, and 5 minutes, shows where mean reversion and momentum dominate. The results vary by pair, time of day, and market conditions.
For a scalper, this informs the choice between fade and follow strategies. If 10-second returns show positive autocorrelation during the London open, momentum entries make sense. If they show negative autocorrelation in the Asian session, fading is preferable.
The numbers are small, often between 0.02 and 0.1, but small correlations can be exploited with enough trades if costs are low.
Compute them across weeks of data, and check stability. If the sign changes randomly, there is nothing to trade.
What do the autocorrelations look like at the horizon the strategy uses?