Distributions of the reconstruction error (RMSE) for a polynomial baseline, a PCA-based baseline, and our method, for different number of fitting parameters. For our model, as well as the PCA baseline, each distribution is the average of three runs using different random seeds controlling the initialization and the train/validation split. The dashed lines indicate the respective medians. Results are shown for the PYATMOS dataset; additional plots for the GOYAL-2020 are found in Fig. D.2 (https://github.com/timothygebhard/ml4ptp/blob/main/scripts/plotting/fig-5-error-distributions/plot-error-distributions.py).