Real-world time series are often shaped by external events and structural shifts, which makes forecasting based only on historical numbers insufficient. Language models can pull in outside news, but standard retrieval-augmented methods struggle with noisy inputs, missing signals, and a lack of causal reasoning about how events affect outcomes.
To address this, researchers propose SEER (Self-Evolving Event Reasoning and Retrieval), a closed-loop framework that dynamically optimizes how events condition time series forecasts. SEER converts prediction errors into two decoupled feedback mechanisms: a reflective retrieval memory that refines later search queries and filters spurious noise, and a persistent causal knowledge base that distills transferable domain dynamics.
The framework enforces strict chronological boundaries across both event retrieval and reflection, which the authors say prevents look-ahead bias and data leakage. Across six volatile time-series benchmarks, SEER consistently outperforms state-of-the-art time series foundation models and language model baselines.