MetaPS: Adaptive Programmatic Strategy Selection for Market Agents

Researchers introduce MetaPS, a framework that enables AI agents to adaptively select from a library of pre-programmed trading strategies based on market conditions, rather than generating actions directly. The system uses market simulations to train models on when to deploy specific strategies, demonstrating consistent improvements across model sizes and outperforming fixed-strategy baselines and direct LLM decision-making approaches.

MetaPS addresses a fundamental challenge in AI-driven trading: no single strategy performs optimally across all market regimes. Traditional approaches either lock agents into fixed strategies or ask language models to generate novel trading actions in real-time, both of which have documented limitations. This research shifts the paradigm toward supervised strategy selection, where an agent learns to recognize market states and dispatch the most appropriate pre-validated program.

The approach builds on established machine learning principles—simulation-based supervision and modular program composition—applied to financial decision-making. By using backtested markets as training data sources, MetaPS generates abundant labeled examples showing which strategies succeed in specific conditions. This grounds abstract market knowledge into concrete state-action patterns that smaller models can learn efficiently.

The implications extend beyond academic interest. Financial institutions constantly balance adaptability with interpretability and risk control. MetaPS offers a framework that maintains both: selected strategies remain auditable code modules rather than opaque neural network outputs, while the selection mechanism adapts to changing conditions. The finding that smaller fine-tuned models (0.8B parameters) outperform larger API-based LLMs suggests that domain-specific training can compensate for parameter disadvantages—a crucial insight for cost-conscious deployment.

The research validates a broader trend toward decomposing complex AI tasks into modular components rather than pursuing end-to-end neural solutions. For trading applications, this decomposition allows institutions to leverage rigorous backtesting infrastructure alongside modern AI techniques. Future work likely explores expanding strategy libraries, real-market validation, and integration with execution infrastructure.

• →MetaPS uses simulation-guided learning to train agents to select from pre-programmed trading strategies rather than generate novel actions.
• →Smaller fine-tuned models consistently outperformed larger API-based LLMs on multi-stock trading tasks, suggesting domain-specific training efficiency.
• →The modular strategy-selection approach maintains interpretability and auditability compared to end-to-end neural decision-making.
• →Market simulations provide scalable supervision for learning adaptive strategy selection across diverse trading environments.
• →Framework demonstrates measurable improvements across model scales from 0.8B to 9B parameters on experimental trading benchmarks.