SEA-TS: Self-Evolving Agent for Autonomous Code Generation of Time Series Forecasting Algorithms

Authors: Longkun Xu, Xiaochun Zhang, Qiantu Tuo, Rui Li Date: 2026-03-05 Paper ID: openalex:2603.04873

Summary

The Self-Evolving Agent for Time Series Algorithms (SEA-TS) framework addresses limitations in conventional ML development by autonomously generating and optimizing forecasting code through an iterative loop. Key innovations include Metric-Advantage Monte Carlo Tree Search (MA-MCTS) for reward-agnostic guidance, automated Code Review with prompt refinement to learn from execution errors, and Global Steerable Reasoning for cross-trajectory knowledge sharing. The approach uses a MAP-Elites archive to maintain architectural diversity. SEA-TS significantly outperforms state-of-the-art methods like TimeMixer, demonstrating its capability to engineer genuinely novel, high-performing algorithmic components for time series forecasting tasks.

Key Contributions

• Proposed SEA-TS, an autonomous framework that generates, validates, and optimizes time series forecasting code via an iterative self-evolution loop.
• Introduced Metric-Advantage Monte Carlo Tree Search (MA-MCTS) for more discriminative guidance in the code search process by using normalized advantage scores instead of fixed rewards.
• Developed a Code Review mechanism with running prompt refinement that encodes corrective patterns from execution errors to prevent the recurrence of similar mistakes.
• Achieved significant performance gains, including a 40% MAE reduction over TimeMixer on the Solar-Energy benchmark, and surpassed human-engineered baselines on proprietary datasets.
• The evolved algorithms discovered novel, physics-informed architectural patterns, such as monotonic decay heads and learnable bias corrections.

Limitations

The performance and robustness of the generated code on domains outside of energy forecasting are not fully explored. The computational cost of the iterative self-evolution loop is likely high.

Open Questions & Future Work

• combining-coding-and-research-agents
• context-pruning-for-cost-reduction
• automated-map-elites-dimension-discovery
• advanced-mcts-search-algorithms
• systematic-domain-knowledge-injection
• multi-objective-optimization-for-deployment

Key Concepts

• Metric-Advantage Monte Carlo Tree Search: A search algorithm variant that uses a normalized advantage score derived from performance metrics instead of fixed rewards to guide exploration in an evolutionary code generation framework.
• Global Steerable Reasoning: A mechanism within an evolutionary agent that compares the current state against the globally best and worst solutions found so far to facilitate cross-trajectory knowledge transfer.

Datasets

• Solar-Energy benchmark

Limitations

The performance and robustness of the generated code on domains outside of energy forecasting are not fully explored. The computational cost of the iterative self-evolution loop is likely high.

Links

• Abstract
• PDF