Why the threshold matters
A strong score can still be a bad trade if the cutoff is too loose. Too low and the model overtrades noise. Too high and the edge never gets used. The threshold is where the score becomes a decision.
How to calibrate it
Start with the score distribution on the validation window. Pick a cutoff that gives you the trade frequency you actually want, then check whether the result still looks sane once costs and slippage are included.
- Lower threshold. More trades, more noise.
- Higher threshold. Fewer trades, cleaner signal.
- Walk-forward recalibration. Keeps the cutoff from going stale.
What to watch for
Score magnitudes drift. A threshold that worked in one regime can become too strict or too loose later. If the trade count swings wildly as the market changes, the cutoff needs to move with it.
The blunt rule
Choose the threshold as a trading decision, not as an afterthought.
Common mistakes
- Picking it after looking at the test set. That is target shopping.
- Leaving it frozen forever. Score distributions do not stay fixed.
- Ignoring frequency. A threshold that trades too often or too rarely is still wrong.