Lab — Move the Threshold

Open the Model Evaluation Lab.

The underlying test cases stay fixed while you move the threshold. That lets you isolate one variable: the policy that converts scores into positive and negative predictions.

Begin at the default threshold. Record the four confusion-matrix counts. Then move the threshold upward until recall clearly drops. Move it downward until false positives clearly increase.

Your evidence note should connect the setting to an outcome:

Raising the threshold from ___ to ___ reduced false positives from ___ to ___, but increased false negatives from ___ to ___. For this fictional use case, I would choose ___ because ___ is the more important error to control.

Do not write “this is the best threshold” unless you define what best means and which evidence supports it.

Evidence to save

• two threshold settings;
• confusion counts for both;
• precision and recall for both;
• the error you prioritized;
• one limitation of the fixed synthetic dataset.