4. Evaluation

We evaluate on three public traces. Two contain a documented regime change; the third is included as a control. All three are downsampled to a common interval before evaluation.

Table 2 reports the mean absolute calibration error for the fixed-window baseline and for the adaptive scheme. The adaptive scheme improves on the baseline by 18% on average across the three traces.

We note that improvement is concentrated in the two traces containing a regime change, and that performance on the control trace is within noise of the baseline, which is the expected behaviour.