A theory used in the context of machine learning where a sequence of predictions is made. The accuracies of past predictions are used to derive both the future prediction and an interval within which the future value will lie with a specified probability. No distributional assumptions are made, but exchangeability (see exchangeable) is assumed.