A trader asked me recently how he could set up his backtest to improve his strategy.
He had added some columns, and while some were modestly predictive, none were as predictive as the Yesterday’s Change from Open % example from one of my strategies that I posted recently.
Here was my suggestion to him:
The thing to focus on is adding features (stat-speak for “columns”) that are highly likely to have predictive value on your original signal in the strategy.
Think about your signal and make sure you’ve got everything about it covered in a column in the backtest.
For example, let’s say your strategy enters long on a gap down today. I’d make sure I had the following data points in the backtest as columns:
- the gap in terms of ATR
- yesterday’s gap in terms of ATR
- yesterday’s change
- yesterday’s close in its range
- yesterday’s volume compared to normal
- today’s open as a position in yesterday’s range
- pre market range
- today’s open as a position in today’s pre market range
etc…
Note that the further away a column is from describing your signal, the less predictive it will be. In our example, the gap from two days ago might be only very slightly predictive, so the improvements you get from it would likely be minor.
Feature selection is hard!
It’s worth thinking deeply about, though, because good, predictive columns pay off quickly in improvements to your strategy.
But more importantly, they provide aha moments that you can use to improve every strategy you create in the future as well.
-Dave
