MultiCast: Zero-Shot Multivariate Time Series Forecasting Using LLMs
Georgios Chatzigeorgakidis, Konstantinos Lentzos, Dimitrios Skoutas
TL;DR
MultiCast addresses the challenge of zero-shot multivariate time series forecasting by encoding multivariate inputs into a single token stream using three dimensional multiplexing schemes, and by applying SAX-based quantization to reduce token usage. The approach enables LLMs to exploit interdimensional correlations without fine-tuning, and is evaluated against traditional and LL-based baselines on real datasets, showing competitive RMSE with notable reductions in inference time when SAX quantization is used. Results highlight a trade-off between token efficiency and predictive accuracy, with performance depending on LLM choice and multiplexing strategy. The work points to broader potential of zero-shot LLMs for time series tasks and CPU-friendly deployment, while suggesting future work on expanding back-ends and applying the method to related problems such as imputation and anomaly detection.
Abstract
Predicting future values in multivariate time series is vital across various domains. This work explores the use of large language models (LLMs) for this task. However, LLMs typically handle one-dimensional data. We introduce MultiCast, a zero-shot LLM-based approach for multivariate time series forecasting. It allows LLMs to receive multivariate time series as input, through three novel token multiplexing solutions that effectively reduce dimensionality while preserving key repetitive patterns. Additionally, a quantization scheme helps LLMs to better learn these patterns, while significantly reducing token use for practical applications. We showcase the performance of our approach in terms of RMSE and execution time against state-of-the-art approaches on three real-world datasets.