Conformal prediction and beyond

Abstract

[en] The role of uncertainty quantification (UQ) has become indispensable with the advent of artificial intelligence and its application to the decision-making. This thesis leverages conformal prediction (CP) as its cornerstone, a pivotal methodology in the field of distribution-free and model-agnostic UQ, which stems from the notion of "conformalizing" predictions to data using the residuals to understand the errors distribution. In particular, in this work some strategies within the CP approach are theoretically justified, and its guarantees and limitations presented. Even though the CP paradigm was classically applied only under "data exchangeability" conditions, this work also reviews some of the most recent and non-trivial efforts to enable CP when this hypothesis is not fulfilled. Lastly, to practically demonstrate CP ability to provide prediction intervals with statistically valid coverage, different strategies are successfully applied both to a tabular data regression problem and to a time series forecasting problem.