Probability Clipping
Probability clipping limits extreme probabilities (near 0 or 1) to avoid infinite penalties in log loss and to reduce the impact of extreme overconfidence.
Definition
Probability clipping is the practice of capping predicted probabilities away from 0 and 1. For example, you might replace p = 1.00 with p = 0.9999 and p = 0.00 with p = 0.0001.
Why it is used
• Log loss becomes undefined at p = 0 or p = 1 (because ln(0) is not defined).
• Clipping reduces the score impact of extreme mistakes driven by overconfidence.
How to choose a clip value
Common clip values are 0.0001 to 0.9999 or 0.001 to 0.999. The right choice depends on sample size and how often you issue very extreme forecasts.
Important caveat
Clipping should be a transparent evaluation setting. It can make scores look better by preventing large penalties. If you publish scorecards, document your clipping rule in methodology.
Related
Clipping is most relevant to log loss, but it also connects to probability discipline and calibration.