Quantifying Predictive Uncertainty Using Belief Functions: Different Approaches and Practical Construction

Abstract

We consider the problem of quantifying prediction uncertainty using the formalism of belief functions. Three requirements for predictive belief functions are reviewed, each one of them inducing a distinct interpretation: compatibility with Bayesian inference, approximation of the true distribution, and frequency calibration. Construction procedures allowing us to build belief functions meeting each of these three requirements are described and illustrated using simple examples.