Fig. 1

Workflow of 10-fold cross-conformal predictor. The training set is randomly permuted and split into ten, non-overlapping folds. An inductive conformal predictor (pink area) is trained for each split, using a single fold as its calibration set and the remaining nine folds as its proper training set. Proper training sets are used for fitting the endpoint and error models. Calibration sets are used to evaluate predictive ability of the model and to accumulate a list of \(\alpha\) (compound nonconformity) values. For any new prediction, each inductive predictor will give an endpoint prediction (single-value prediction) and produce a prediction interval based on the predicted error, the desired confidence and the list of \(\alpha\) values. The final prediction is computed by aggregating the individual predictions using the median midpoint and median interval width