Title: Additive quantile regression for electricity demand forecasting
Authors: Matteo Fasiolo - University of Bristol (United Kingdom) [presenting]
Abstract: Generalized Additive Models (GAMs) are flexible and interpretable statistical models that are widely used in applied statistics, especially since the advent of efficient and stable methods for smoothing parameter selection. We will describe a computationally efficient and well-calibrated framework for extending GAM methodology to quantile regression models. The proposed methods are based on the general belief updating framework to loss based inference, but they are computed by adapting the stable fitting methods. Fast computation is enabled by adopting a smooth generalization to the quantile regression pinball loss which, if tuned correctly, can also be statistically superior to the original loss. The belief updating framework requires selecting a learning rate balancing the loss with the prior during inference, hence we will present a novel calibration method for selecting this parameter, which aims at obtaining reliable quantile uncertainty estimates. We will briefly discuss the implementation of the proposed methods in the qgam R package, and we will illustrate their performance in the context of an electricity demand forecasting application.