Johan Hallberg Szabadváry: Modified conformal prediction for time series forecasting

Abstract

In this talk we start with a brief introduction to conformal prediction (CP) which is today widely used as a ML algorithm. Briefly speaking, CP uses past experience to determine the level of confidence in new predictions. In the on-line setting, where examples arrive successively, CP together with some method that produces a prediction of a label constructs a prediction set. The main assumption of the classical CP is that data is exchangebale, meaning essentially that any permutation of the data is equally probable. However, data is rarely exchangeable (or exactly agree with whatever assumption we make), and violations of the assumption leads to loss of validity. We will discuss some modifications in this case.