CaReTS: A Multi-Task Framework Unifying Classification and Regression for Time Series Forecasting

TL;DR

CaReTS tackles multi-step time series forecasting by explicitly disentangling macro-level trends from micro-level deviations through a dual-stream, multi-task framework. It combines trend classification and deviation regression with an uncertainty-aware loss to balance three objectives, and supports four variants across CNN/LSTM/Transformer encoders. Empirical results on electricity price and unmet power forecasting show state-of-the-art RMSE and high trend accuracy, with the Transformer-based CaReTS2 achieving the best overall performance and a favorable compute profile. The approach enhances interpretability while maintaining accuracy, and includes open-source reproducibility for broader adoption.

Abstract

Recent advances in deep forecasting models have achieved remarkable performance, yet most approaches still struggle to provide both accurate predictions and interpretable insights into temporal dynamics. This paper proposes CaReTS, a novel multi-task learning framework that combines classification and regression tasks for multi-step time series forecasting problems. The framework adopts a dual-stream architecture, where a classification branch learns the stepwise trend into the future, while a regression branch estimates the corresponding deviations from the latest observation of the target variable. The dual-stream design provides more interpretable predictions by disentangling macro-level trends from micro-level deviations in the target variable. To enable effective learning in output prediction, deviation estimation, and trend classification, we design a multi-task loss with uncertainty-aware weighting to adaptively balance the contribution of each task. Furthermore, four variants (CaReTS1--4) are instantiated under this framework to incorporate mainstream temporal modelling encoders, including convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and Transformers. Experiments on real-world datasets demonstrate that CaReTS outperforms state-of-the-art (SOTA) algorithms in forecasting accuracy, while achieving higher trend classification performance.

Figures (9)

• Figure 1: Two types of dual-stream CaReTS architectures
• Figure 2: RMSE on power across approaches
• Figure 3: RMSE on price across approaches
• Figure 4: RMSE across forecasting steps using CaReTS2-Transformer
• Figure 5: Trend accuracy across forecasting steps using CaReTS2-Transformer