Time-series forecasting is an important task in many domains, including finance, weather prediction, and energy consumption forecasting, and deep learning methods have emerged as the best-performing time-series forecasting methods over the last few years. However, most proposed time-series forecasting models are deterministic and are prone to errors when deployed in production, potentially causing significant losses and penalties when making predictions with low confidence. In this paper, we propose the Time-Energy Model (TEM), a framework that introduces so-called selective time-series forecasting using energy-based models (EBMs). Selective forecasting estimates model confidence and allows the end-user to selectively reject forecasts while maintaining a desired target coverage. TEM is model-agnostic and can be used to improve forecasting accuracy of any encoder-decoder deterministic time-series forecasting model. TEM is trained using a combi-nation of supervised and self-supervised learning, leveraging excellent single-point prediction accuracy while maintaining the ability to reject forecasts based on model confidence. Experimental results indicate that TEM generalizes well across 5 state-of-the-art deterministic time-series forecasting models and 5 benchmark time-series forecasting datasets. Using selective forecasting, TEM reduces prediction error by up to 49.1% over 5 state-of-the-art deterministic models. Furthermore, TEM has up to 87.0% lower error than selected baseline EBM models, and achieves significantly better performance than state-of-the-art selective deep learning models. Code for the proposed TEM framework is available at.
Brusokas et al. (Wed,) studied this question.