Title: Byte Pair Encoding for Efficient Time Series Forecasting

Abstract: Current time series tokenization techniques typically map a fixed number of data points to single tokens. This rigid methodology often leads to token redundancy, particularly when handling simple structures such as prolonged constant values, thereby imposing significant computational burdens. Drawing inspiration from the efficacy of byte pair encoding, we present the inaugural pattern-focused tokenization framework for time series analysis. By leveraging a discrete vocabulary of common motifs, our approach consolidates data samples exhibiting underlying patterns into tokens, enabling adaptive compression of the time series. Furthermore, we introduce conditional decoding—a lightweight, post-hoc optimization technique that capitalizes on the discrete motif set and the continuous nature of time series data. This method enhances performance without requiring gradient calculations or incurring additional computational costs. Evaluations on contemporary time series foundation models reveal that our motif-driven tokenization enhances forecasting accuracy by an average of 40% and increases efficiency by 2314%. Additionally, conditional decoding achieves up to a 48% reduction in Mean Squared Error (MSE). Our comprehensive analysis highlights the tokenization’s adaptability across varied temporal patterns, its ability to generalize to unseen data, and the semantic richness of its tokens, which effectively capture distinct time series attributes such as trends and statistical moments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC