Robust sequential learning with applications to the forecasting of air quality and of electricity consumption

Abstract

In this talk I will describe a setting of sequential learning in which many expert forecasts are available at each round to predict quantitative outcome. The question is how to combine these expert forecasts. We will do so in a robust manner, not assuming any stochasticity of the sequence of outcomes to predict. First, algorithms and performance bounds will be presented. Then, applications to real data sets will be discussed: one is concerned with air quality and the other one with electricity consumption.