Improve searches for a parameter variant that does better than your current strategy on out-of-sample data, and only lets you save one that actually does.
Why this is different from Optimize
Optimize shows you the best fitted result - the number computed on the same history the parameters were chosen against - and lets you save it. That is how strategies get overfit: the best of a big search is partly luck fit to that history's noise, and it regresses forward.
Improve judges every variant on its out-of-sample expectancy - its performance on the recent slice the search held back and never optimized against - and compares it to your current strategy on that same window. A variant can only look good there if it has a real edge on data it was not fit to. So a saved Improve variant is far more trustworthy than an Optimize winner.
How to read the result
- Out-of-sample per trade is the honest forward number. Judge on this, not on the fitted total return.
- vs current is the improvement over what you already have. Only variants that beat your current strategy are offered to save.
- If nothing beats your current strategy, the limit is the strategy's structure (its logic), not its parameters. More tuning will not help - change the logic instead: add a filter, use a different signal, or reduce the degrees of freedom.
What "out-of-sample" means
Out-of-sample is data the parameters were never tuned on. A backtest on the data you fit to will almost always look good - it has seen the answers. Holding back a recent window and measuring there is the closest a backtest gets to "what happens next", which is the only number worth trusting before you risk real money.