_Sort

The "optimal" parameters were sitting on a cliff The single best point in parameter space is often a local maximum that's extremely fragile. Tiny changes in environment — wider spreads, slight latency — and you fall off.

Regime filters mostly have too much delay to actually act upon them, at least what I found. To avoid over fitting you should keep part of the test set outside of the optimisation process so you have something clean to run on. I run on crypto so usually I ran the optimised version on a different asset (hopefully somewhat "uncorrelated").

Conformal prediction research on conformal prediction which is about making intervals of prediction using previous prices. The main thing is that whatever model is used, it is possible to predict a price range, on average, with a chosen accuracy. In other words, one can use any way to predict the next price, and end up with ranges where the price will be in X % of the time on average, where X is chosen.

Don't be fooled by rising and falling tides of the market Trading patterns of last 20 days Absent of indicator is also an indicator Clinging effect (holding on in defiance of logic) Adrenaline effect (buying in defiance of logic)