WorksMathematics
Time-Varying Recalibration of Forecast Distributions
Abstract
Combined density forecasts are often miscalibrated, and the miscalibration can change over time. We propose a time-varying framework that recalibrates any combined forecast through applying a monotone map to its probability integral transform. We consider a Beta CDF or a rational quadratic spline as the map, and their parameters are updated either via score-driven filters combined with dynamic model averaging, or by a convolution-based update that additionally conditions on features of the forecast distribution. In simulations, the spline map corrects multimodal miscalibration that the Beta map cannot, and the convolution update achieves the highest log score when the miscalibration switches between regimes. We apply the method to nine G10 exchange rates and US inflation, recalibrating forecast distributions from both a loss-discounting linear pool and Bayesian predictive synthesis. In both applications, recalibration improves PIT uniformity, with gains in forecast accuracy most evident when the original forecasts are miscalibrated.