DualRes: A Resampling-Based Framework for Enhancing Probabilistic Forecasting

Tingting Wang, Yunyi Zhang, Benyou Wang

Keywords: probabilistic forecasting, Time Series, Resampling, Multivariate Forecasting

Abstract: Probabilistic forecasting of time series has gained increasing attention in practice due to the need for assessing risks and uncertainties in future observations. In this manuscript, we propose DualRes, a framework that improves the probabilistic forecasting performance of existing algorithms by incorporating conditional heteroskedasticity and residual distributional information. Specifically, during training, DualRes employs two separate models to learn the conditional mean and volatility of the time series, while during inference it generates pseudo-normalized residuals through resampling. DualRes requires only mean forecasts, so it offers substantial flexibility in the choice of forecasting algorithms-even algorithms originally designed for mean forecasting can be adapted to probabilistic forecasting. DualRes applies to both univariate and multivariate time series and remains robust under non-Gaussian errors with conditional heteroskedasticity. Numerical experiments on six real-world datasets demonstrate its good empirical performance in capturing distribution of future observations and producing accurate prediction intervals.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 12580