Optimizing parameters on a year of one-minute bars is easy, and the results look beautiful. A moving average period of 13, a threshold of 7.5, a stop of 4.2 pips. All tuned to the past.
The problem is that minute data has an enormous number of small moves, so with enough tuning knobs, something will fit. It does not mean it will work next month.
Signs of overfitting include very specific parameter values, a sharp drop in performance when a parameter changes slightly, and results that depend on a handful of trades.
A healthier approach is testing parameter stability. If the strategy works with a period of 10, 13, and 16, the idea may be sound. If it only works at 13, it is probably noise.
Out-of-sample testing is the other basic tool. Reserve a period never used in optimization and test once. A big deterioration is a warning.
Another approach is keeping the number of parameters small, since each one adds room for self-deception.
How is overfitting checked, and has a beautiful backtest ever fallen apart in live trading?
Advertisement
Overfitting on one-minute data
Recommended broker for automated trading & scalping
