Title: From Feature Space to Autoregression: A Novel Strategy for Zero-Shot Time Series Forecasting

Abstract: Zero-shot time series forecasting seeks to predict future values for series that the model has never encountered before, necessitating the ability to generalize temporal dynamics beyond the scope of the training distribution. Although recent foundation models have demonstrated impressive performance within their training domains by leveraging large-scale pretraining, their success is frequently contingent upon extensive data coverage and the implicit memorization of patterns. This reliance can hinder generalization capabilities, particularly when data is limited or when there is a significant divergence between source and target domains. To address these challenges, we introduce FSA, a framework that maps features to strategies for controlled zero-shot univariate forecasting. Rather than processing raw sequences directly in the observation space, FSA establishes a structured mapping from an interpretable feature space to an autoregressive strategy space. This architectural choice embeds explicit inductive biases that separate global trends, periodic elements, and local temporal dynamics, allowing the model to identify transferable time-series structures with reduced reliance on data assumptions. Our empirical evaluations indicate that, when compared against Transformer-based architectures under identical pretraining data, training protocols, and similar parameter constraints, FSA achieves superior performance in our controlled zero-shot setting.

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