Alpha-R1: Alpha Screening with LLM Reasoning via Reinforcement Learning

Battle Count: 163

Directly addresses a core challenge in quantitative trading: dynamic factor screening in non-stationary markets. Proposes a novel semantic gating mechanism that bridges static factor definitions and dynamic market regimes. Demonstrates strong out-of-sample performance (47.87% AR on S&P 500, 40.57% on CSI 300) with rigorous statistical validation including block-bootstrap inference, risk attribution, and falsification tests. The second-stage reranking approach is practically deployable within existing quantitative workflows. However, capacity constraints and single-year validation limit immediate production deployment.

Requires multi-stage pipeline: (1) LLM-based data abstraction and memory construction, (2) factor backtesting and semantic profiling, (3) GRPO reinforcement learning training on 64 H800 GPUs (~120 hours), (4) portfolio construction with slot rotation and VWAP execution. Multiple LLMs needed (Qwen3-8B for reasoning, Claude 3.7 Sonnet for state synthesis/profiling, Claude 3.5 Haiku for judging). Complex reward function with market feedback, LLM-as-judge consistency penalty, and structural penalties. Inference is simpler (single 8B model, ~15 seconds per decision). Requires access to proprietary data feeds and LLM APIs.

Full implementation and resources available at GitHub (https://github.com/FinStep-AI/Alpha-R1). Detailed appendices include prompts, hyperparameters, data samples, and execution protocol. However, proprietary market data feeds (Tushare, Massive), news sources (Sina Finance), and LLM API access (Claude 3.7 Sonnet, Claude 3.5 Haiku) are required for exact reproduction. Training requires 64 H800 GPUs for ~120 hours. Anti-leakage protocol and temporal splits are clearly documented.