Fig. 9From: Uncertain of uncertainties? A comparison of uncertainty quantification metrics for chemical data setsa Error-based calibration plot for the epistemic evidential uncertainties one of the vertical IP models. b Distribution of errors according to their Z-value for the model split with random seed = 42 compared with a Gaussian distribution of width 1. Error-bars are Poisson.We see that the non-zero miscalibration area is mostly caused by an increased number of low errors compared to what is expected from the uncertainty-distribution. This is also clear from the error-based calibration plot, where we observe that it is the errors for the high uncertainties that are significantly lower than what is expected from the corresponding uncertainties, i.e. these uncertainties are overestimatedBack to article page