A Novel Smoothed Loss and Penalty Function for Noncrossing Composite Quantile Estimation via Deep Neural Networks
The episode discusses a paper by Hatalis et al. that introduces a novel smoothed loss and penalty function for noncrossing composite quantile estimation using Deep Neural Networks to improve wind power forecasting. The authors solve the quantile crossover problem by using a logistic-based smooth approximation of the pinball loss and adding a penalty term to enforce correct ordering, showing better performance than existing methods.