Predicting the Future by Retrieving the Past

2025-11-08

AI Summary

Existing univariate time series forecasting models—such as MLPs, Transformers, and TCNs—rely solely on fixed-length local historical windows, limiting their ability to explicitly capture and leverage global temporal patterns, thereby constraining predictive accuracy. To address this, we propose MemTS: a framework that constructs a scalable global memory bank to store historical time series segments and introduces a similarity-based dynamic retrieval mechanism to adaptively identify and extract the most relevant historical patterns for the current query. These retrieved global patterns are then fused with predictions from any off-the-shelf local forecasting model to produce the final output. MemTS is the first method to enable explicit, interpretable, and plug-and-play enhancement of forecasting models with global historical knowledge. Extensive experiments across seven real-world benchmarks demonstrate that MemTS consistently improves state-of-the-art models by an average of 8.4%, significantly outperforming existing approaches.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Planning, Routing, and Scheduling: Temporal Planning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Deep learning models such as MLP, Transformer, and TCN have achieved remarkable success in univariate time series forecasting, typically relying on sliding window samples from historical data for training. However, while these models implicitly compress historical information into their parameters during training, they are unable to explicitly and dynamically access this global knowledge during inference, relying only on the local context within the lookback window. This results in an underutilization of rich patterns from the global history. To bridge this gap, we propose Predicting the Future by Retrieving the Past (PFRP), a novel approach that explicitly integrates global historical data to enhance forecasting accuracy. Specifically, we construct a Global Memory Bank (GMB) to effectively store and manage global historical patterns. A retrieval mechanism is then employed to extract similar patterns from the GMB, enabling the generation of global predictions. By adaptively combining these global predictions with the outputs of any local prediction model, PFRP produces more accurate and interpretable forecasts. Extensive experiments conducted on seven real-world datasets demonstrate that PFRP significantly enhances the average performance of advanced univariate forecasting models by 8.4%. Codes can be found in https://github.com/ddz16/PFRP.

Problem

• Existing models cannot dynamically access global historical patterns during inference
• Local forecasting models underutilize rich patterns from global time series history
• Current approaches lack explicit integration of global knowledge for time series prediction

Innovation

• Global Memory Bank stores historical patterns
• Retrieval mechanism extracts similar past patterns
• Combines global predictions with local models