A decision threshold is a cutoff that turns a model’s score or estimated probability into a predicted class or action. If the estimated probability is p and the cutoff is τ, a simple binary rule predicts positive when p ≥ τ and negative otherwise. The model produces the probability; the threshold applies a decision rule to it. section 11.3.1 section 3.3.2
The choice of cutoff shapes the decisions made from the same probability estimates:
- Lowering the threshold can reduce missed positive cases while increasing false alarms; the appropriate balance depends on the relative costs of those errors. section 11.3.1
- A threshold is distinct from model training: the model learns to estimate probabilities, while a predetermined cutoff converts those estimates into decisions. section 11.3.1
Pedestrian detection
A vehicle estimates the probability that a pedestrian is present and brakes when that probability is at least 0.3. Choosing a cutoff below 0.5 reflects the greater cost assigned to failing to detect a pedestrian than to braking unnecessarily. section 11.3.1 section 11.3



















