Fig. 4. Performance evaluation for regression when the underlying model is known.

a Mean absolute error (MAE) and b expected normalised calibration error (ENCE)⁸² achieved by the Multi-SWAG models trained on only one model, plotted for different trajectory lengths T by averaging over 5 × 10⁴ (FBM, SBM) or 4 × 10⁴ (ATTM, LW, CTRW) test trajectories each. The ENCE characterises the mean difference between predicted and observed errors. As was the case for the unknown ground truth model (Fig. 3), we can achieve a small calibration error below 3%. The MAE shows the expected results with regards to the histograms in Fig. 5.