Overfitting is easier on scalping systems because the sample looks huge
Posted: Sat Oct 31, 2026 7:55 pm
Thousands of trades in a backtest feel like strong evidence. For a scalping system, they're often less informative than they appear.
Trades clustered in the same hours, under the same volatility regime, aren't independent. Parameter tuning on that data can fit noise: entry thresholds, filter values and exit rules chosen to maximise past results.
The more parameters, the easier the fit. A system with eight tuned numbers can look brilliant on its training period and fall apart afterward.
Signs of overfitting include sharp performance peaks at specific parameter values, results that collapse when a setting changes slightly, and strategy logic that can't be explained simply.
A better approach is to look for plateaus, regions where nearby parameter values all work. It also helps to test on different instruments, different years and different market conditions.
If a system only works on one pair during one year with exact settings, it's probably a coincidence.
A practical guardrail is to reserve the most recent year of data and never touch it until the design is frozen. Looking at it once, honestly, tells more about the strategy's real quality than another month of tuning on the same old sample.
How many parameters does your strategy have, and how do you test their stability?
Trades clustered in the same hours, under the same volatility regime, aren't independent. Parameter tuning on that data can fit noise: entry thresholds, filter values and exit rules chosen to maximise past results.
The more parameters, the easier the fit. A system with eight tuned numbers can look brilliant on its training period and fall apart afterward.
Signs of overfitting include sharp performance peaks at specific parameter values, results that collapse when a setting changes slightly, and strategy logic that can't be explained simply.
A better approach is to look for plateaus, regions where nearby parameter values all work. It also helps to test on different instruments, different years and different market conditions.
If a system only works on one pair during one year with exact settings, it's probably a coincidence.
A practical guardrail is to reserve the most recent year of data and never touch it until the design is frozen. Looking at it once, honestly, tells more about the strategy's real quality than another month of tuning on the same old sample.
How many parameters does your strategy have, and how do you test their stability?