Fast additive quantile regression for electricity demand forecasting
Matteo Fasiolo (University of Bristol)
Friday 2nd November, 2018 15:00-16:00 Maths 311B
Generalized Additive Models (GAMs) are an extension of Generalized Linear Models (GLMs), which
allow the inclusion of random effects, complex non-linear effects (built using spline bases) and, more recently, response distributions outside the exponential family. In this talk I will describe a computationally efficient Bayesian framework for fitting quantile GAMs, which are based on the pinball loss, rather than on a probabilistic response distribution. The new hierarchical framework selects both the smoothing parameter and the so-called "learning-rate" automatically and efficiently, and provides posterior credible intervals that are approximately calibrated in a frequentist sense. I will illustrate the new methods in the context of electricity demand forecasting, where they provide a predictive performance that is competitive with that of Gradient Boosting (GB), but at a fraction of GB's computational cost.