SEER: Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting

SEER. Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting. By Mingtian Tan and eight others, from Google Cloud AI Research and the University of Virginia. Published on the second of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Real-world time series are often moved by things outside the series: exogenous events and structural shifts. Forecasting from historical numbers alone cannot see them. Language models can retrieve external news, but standard retrieval-augmented approaches run into three problems. The retrieved material is noisy. Relevant signals go missing. And there is no ability to reason causally about what an event will do to the series.

SEER, short for Self-Evolving Event Reasoning and Retrieval, is a closed-loop framework that dynamically optimizes how a forecast is conditioned on events. The forecaster's own prediction errors are what drive the loop. When a forecast misses, the miss is not simply recorded. It becomes feedback that changes what the system retrieves and what it believes about how events act on the series.

Those errors are translated into two feedback mechanisms, kept decoupled from each other. One is a reflective retrieval memory, which refines later search queries and filters out spurious noise. The other is a persistent causal knowledge base, which distills domain dynamics that transfer. Over both, SEER enforces strict chronological boundaries, for event retrieval and for reflection alike. That is what 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 as well. Those are the two families the framework sits between: models that read only the numbers, and models that read the news.

Two memories, then: one that sharpens the search and one that accumulates cause and effect, both learning from mistakes and both walled off from the future. Without that wall, a system that learns from its own errors could quietly learn from tomorrow's news instead.