Latency in the tester versus latency in reality
Posted: Tue Nov 03, 2026 12:03 am
Latency in the tester versus latency in reality
Backtests typically assume orders fill instantly at the signal price. In reality, there is a delay between decision and fill, and in that delay price moves. For a scalper with a tight target, even 200 milliseconds can matter.
The cost of latency depends on volatility. In a quiet tape, price barely moves in a fraction of a second. During active periods, which is when many scalping signals fire, price can move a pip or more in the delay.
To model it, add a delay to the simulated fills, using the observed distribution of real latency, not a single average. Apply the first available price after the delay, using the ask for buys and the bid for sells. Then compare the results with the zero-latency case.
The difference is often large, and it widens during volatile hours. A system that only works with zero latency is not tradable.
If the tester does not allow latency, post-process the signals with a custom script using tick data.
How much did performance drop after adding a realistic latency model to the backtest?
Backtests typically assume orders fill instantly at the signal price. In reality, there is a delay between decision and fill, and in that delay price moves. For a scalper with a tight target, even 200 milliseconds can matter.
The cost of latency depends on volatility. In a quiet tape, price barely moves in a fraction of a second. During active periods, which is when many scalping signals fire, price can move a pip or more in the delay.
To model it, add a delay to the simulated fills, using the observed distribution of real latency, not a single average. Apply the first available price after the delay, using the ask for buys and the bid for sells. Then compare the results with the zero-latency case.
The difference is often large, and it widens during volatile hours. A system that only works with zero latency is not tradable.
If the tester does not allow latency, post-process the signals with a custom script using tick data.
How much did performance drop after adding a realistic latency model to the backtest?