A calibration curve (also known as a reliability diagram) is a graphical tool to assess how well a probabilistic classification model is calibrated​.It plots the actual observed frequency of the positive class against the predicted probability, typically by partitioning the predictions into bins. For example, one might bin all instances where the model predicted between 0.7 and 0.8 probability, and then compute the fraction of positives in that bin​. On the calibration plot, the x-axis is the average predicted probability in the bin and the y-axis is the true fraction of positives in that bin. An ideal calibration curve would lie on the diagonal line $y = x$, meaning the model’s predicted probability equals the true outcome If the curve falls below the diagonal, the model is over-predicting (outcomes occur less often than predicted); if it’s above the diagonal, the model is under-predicting. Calibration curves thus help diagnose if a model’s confidence estimates need adjustment. They are often presented alongside classification results to complement metrics like accuracy – a model might be accurate but still poorly calibrated, which a calibration curve would reveal.