Classical ML, validation, metrics, interpretation, debugging · insight-0142

Treat a negative test R² as a signal of shift or broken features, not as a rounding curiosity

Situation
A regression model that scored well in validation shows negative R² on a new test period or segment.
Action
Compare against the constant baseline, check for level shifts, broken or missing features and target definition changes, and inspect residuals over time before tuning anything.
Reason
A negative R² means squared errors exceed the variance around the test mean, which usually reflects a systematic offset or a pipeline fault.
Boundary
Out-of-time tests, new markets, post-incident data and model monitoring.

Symptom or risk

The model is worse than predicting the test set's own mean; decisions based on it are worse than a constant.

Very small test sets with tiny variance can give negative R² from noise; check the absolute errors.

A quick recalibration of the level may restore value while the root cause is investigated.

Clipping R² at zero in dashboards, which hides the problem.

R² can be negative when the model is arbitrarily worse than a constant model that predicts the mean, which scores 0.
Metrics and scoring: quantifying the quality of predictions (scikit-learn User Guide 3.4)