Feature to Dynamics: Feature-space to Autoregression strategy for Zero-shot Time Series Forecasting

2026-05-31

AI Summary

This work addresses the challenge of weak generalization in zero-shot time series forecasting, particularly when training and target domains do not overlap or when data is scarce. To this end, the authors propose the Feature-to-Strategy Autoregression (FSA) framework, which introduces explicit inductive biases by mapping an interpretable feature space to an autoregressive policy space. This approach explicitly decouples global trends, periodicity, and local dynamics, enabling structured univariate zero-shot prediction. In contrast to conventional methods that model directly in the observation space, FSA leverages feature modeling and policy generation mechanisms, substantially reducing reliance on extensive data coverage and pattern memorization. Under identical pretraining data, training protocols, and model scales, FSA outperforms existing Transformer-based architectures in controlled zero-shot forecasting settings.

Abstract

Zero-shot time series forecasting aims to predict future values for previously unseen series, requiring models to generalize temporal dynamics beyond the training distribution. While recent foundation models achieve strong in-domain performance through large-scale pretraining, their effectiveness often relies on broad data coverage and implicit pattern memorization, which can limit generalization when data are scarce or source and target domains are disjoint. In this work, we propose FSA, a feature-to-strategy framework for controlled zero-shot univariate forecasting. Instead of directly modeling raw sequences in the observation space, FSA learns a structured mapping from an interpretable feature space to an autoregressive strategy space. This design introduces explicit inductive biases that disentangle global trends, periodic components, and local temporal dynamics, enabling the model to capture transferable time-series structure with fewer data assumptions. Empirical results show that, under identical pretraining data, training protocol, and comparable parameter budgets, FSA outperforms Transformer-based architectures in our controlled zero-shot setting.

Problem

• zero-shot time series forecasting
• domain generalization
• temporal dynamics
• data scarcity
• unseen series

Innovation

• zero-shot forecasting
• feature-to-dynamics
• autoregressive strategy
• inductive bias
• time series representation