Work out model drift score instantly with clear inputs, formula shown and shareable results.
Prediction drift is monitored by comparing the current output distribution against a stored baseline. Expressing the mean shift in baseline standard deviations makes the comparison scale-free and directly usable as an alert threshold: two standard deviations flags roughly the most extreme 5 percent of normal variation, three flags 0.3 percent. This catches upstream data changes before labels arrive to reveal accuracy loss.
Standardised drift
z = (current mean - baseline mean) / baseline standard deviation; alert when |z| >= threshold
Labels usually arrive days or weeks later. The prediction distribution is available immediately, so it is the earliest available signal that inputs have changed.
No. It can reflect a genuine change in the population, such as a seasonal shift. Drift is a trigger to investigate, not proof that the model has degraded.