For many applications it is important to quantify the accurracy of the predictions produced by a model. This accurracy can be categorized in two ways: the accurracy of the prediction of the true regression and that of the prediction with respect to the observed output. We will refer to the latter through the prediction intervals.
Figure 1: Procedure used for estimating the prediction intervals.
We show in Figure 1 the procedure used for estimating the prediction intervals.
Figure 2: Prediction intervals (grey area) and expected target values (dashed
line) for the noisy Mackey Glass series.
Figure 2 shows the 95% prediction intervals and the target values for the last 100 points of the test set, this interval is giving the region of the sampled predictive ditribution with smaller error. Note that the prediction intervals of the Bayesian learning fall short of the real value in the peak areas.