Mixing Probabilities, Priors and Kernels via Entropy Pooling

Abstract

We show how to mix machine learning signals such as kernel smoothing and fuzzy memberships via the Entropy Pooling approach by Meucci (2008). We illustrate a case study, where we overlay an exponentially time-decayed prior to a pseudo-Gaussian kernel that emphasizes market scenarios where volatilities and interest rates are similar to today’s levels. The code for the case study is available for download.

Keywords: Black-Litterman, effective number of scenarios, fuzzy membership, Mahalanobis distance, pseudo-Gaussian kernel, crisp conditioning, exponential decay, machine learning

JEL Classification: C1, G11