The decision threshold lab

A small change with visible consequences

A model produces a score. A threshold turns that score into a decision. Use this interactive example to separate those two steps and see how the same scores lead to different errors.

Three things to try

1. Start at 0.50. Record accuracy, recall, and weighted error cost.
2. Move to 0.30. Which mistakes disappeared? Which became more common?
3. Keep the threshold fixed and change the cost of a missed positive. Which numbers change?

What the example does and doesn’t show

The labels and scores are synthetic. The selected thresholds illustrate a trade-off; they are not an estimated optimum. In a real project, choose a threshold using validation data and the decision context, then report its performance on a held-out test set.

Changing the cost does not retrain the model or automatically choose a threshold. It changes the way the errors are valued.