Bootstrapped Exploration with Causal Reasoning: A Training Paradigm for Adaptive Forecasting Agent

Qingwen Zeng, Dajun Guo, Zhaoge Bi, lining chen, Jushang Qiu, Yitian Yang, Carl Yang, Huaming Chen, Ling Chen

Keywords: Time Series Forecasting, Agent, LLM, Agent Training Paradigm

Abstract: Time series forecasting is critical in domains such as finance, energy, and healthcare, yet real-world datasets often exhibit non-stationarity, noise, missing values, and distribution shifts, posing severe challenges for generalization. In practice, industry solutions typically rely on customized forecasting frameworks that combine imputation, decomposition, and specialized models. However, such frameworks incur high labor costs. Moreover, we observe that many frameworks suffer from the impacts of distribution shifts, which degrade their respective performance. Thus, it is critical to establish a new paradigm that retains high transferability across diverse datasets while accumulating reusable strategy knowledge. This is fundamental for large-scale and dynamic environments. While large language model-based agents have recently demonstrated strong reasoning and tool-use capabilities, no forecasting agent can automatically adapt to diverse time-series datasets. This gap arises from two core obstacles: the scarcity of labeled supervision and the inherent complexity of mapping dataset-specific meta-features to effective forecasting strategies. To address these challenges, we propose BECRA, a novel agent training paradigm that learns forecasting intelligence through contrast-aware exploration and causal lesson extraction, without any human-annotated supervision. BECRA distills symbolic strategy lessons that enable in-context planning on unseen datasets, achieving zero training adaptation.

Primary Area: learning on time series and dynamical systems

Submission Number: 6768