Time Series Embedding and Combination of Forecasts: A Reinforcement Learning Approach

2025-08-28

AI Summary

Traditional ensemble methods—particularly simple averaging—struggle to consistently outperform individual base models in time series forecasting, facing a fundamental “prediction combination dilemma.” Method: This paper proposes a reinforcement learning (RL)-based dynamic model selection and weighted ensemble framework. It employs temporal embeddings to capture the dynamic dependence structure of forecast errors and leverages an RL policy to adaptively learn optimal, time- and uncertainty-aware weights for each base model—thereby overcoming the limitations of static weighting schemes. Contribution/Results: Extensive experiments on the M4 competition dataset and the Federal Reserve’s Survey of Professional Forecasters (SPF) demonstrate that our approach significantly outperforms simple averaging, fixed-weight ensembles, and state-of-the-art ensemble baselines. It exhibits superior robustness and generalization across multiple forecasting horizons. The framework establishes a novel, learnable, interpretable, and scalable paradigm for time series forecast fusion.

Technology Category

• Machine Learning: Ensemble Methods
• Reasoning under Uncertainty: Sequential Decision Making
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

The forecasting combination puzzle is a well-known phenomenon in forecasting literature, stressing the challenge of outperforming the simple average when aggregating forecasts from diverse methods. This study proposes a Reinforcement Learning - based framework as a dynamic model selection approach to address this puzzle. Our framework is evaluated through extensive forecasting exercises using simulated and real data. Specifically, we analyze the M4 Competition dataset and the Survey of Professional Forecasters (SPF). This research introduces an adaptable methodology for selecting and combining forecasts under uncertainty, offering a promising advancement in resolving the forecasting combination puzzle.

Problem

• Addressing the forecasting combination puzzle of outperforming simple averages
• Developing dynamic model selection using Reinforcement Learning framework
• Selecting and combining forecasts under uncertainty conditions

Innovation

• Reinforcement Learning dynamic model selection
• Time Series Embedding combination framework
• Adaptable methodology under uncertainty