AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

Battle Count: 51

Directly addresses the core problem of autonomous alpha discovery for quantitative trading. The system produces platform-validated equity alphas with Sharpe ratios up to 3.48 and Fitness up to 9.50 on WorldQuant BRAIN. It introduces novel architectural principles (artifact-level search, adversarial verification with veto authority, pending-aware budget allocation) that address fundamental challenges in LLM-driven factor mining. The evaluation uses a production trading platform with economic stakes, making results directly relevant to practitioners. The paper also addresses the reproducibility crisis in LLM-based trading research.

Moderate complexity. The core discovery logic is remarkably compact (101 lines of role prompts), but the full system requires: (1) LLM API integration with multiple backends, (2) MCTS scheduler with progressive widening, pending-aware UCB, backpressure, and root fallback, (3) WorldQuant BRAIN platform API integration for simulation and submission, (4) halving tournament orchestration (16→8→4→2→1), (5) adversarial reviewer with re-execution capability, (6) artifact freezing and evidence management, (7) parallel worker coordination (validated at 10 concurrent workers). The deterministic code handles only platform I/O and tree bookkeeping; all research logic lives in prompts.

Exceptional reproducibility. The paper releases the complete prompt-to-factor trail: all proposer/reviewer/dispatcher prompts (101 lines), prompt history, scheduler code, complete MCTS state, all candidate reports, simulation inputs and responses, rankings, correlation records, submission checks, final submission responses, and literal executable FASTEXPR strings for all 60 submissions. This is described as the first complete prompt-to-factor release in the audited LLM trading literature. Evaluation uses WorldQuant BRAIN as an external, independently governed platform where authors cannot alter data, simulator, metrics, gates, or grades.