Cross-Stock Predictability via LLM-Augmented Semantic Networks

Battle Count: 78

Highly relevant to quantitative trading. The paper directly addresses cross-stock predictability, a core alpha source in systematic trading. It proposes a concrete, backtested long-short equity strategy with clear signal construction (mean-reversion z-scores aggregated via relation-aware softmax weights). The LLM-based edge filtering provides a novel way to improve network fidelity for pairs-trading-like strategies. Results show statistically significant Sharpe ratio improvement (0.742 to 0.820) and reduced drawdown. The framework is practical for institutional equity strategies, though the reliance on annual 10-K data limits intraday applicability. The parameter sensitivity analysis and ablation studies provide actionable guidance for implementation.

Moderate complexity. Stage 1 (embedding similarity + top-K graph) is straightforward with standard NLP and graph libraries. Stage 2 (LLM edge classification) requires careful prompt engineering, API integration, batching, caching, and JSON parsing, but is well-specified with the provided prompt template. The signal construction (Gatev distance, z-scores, per-stock softmax) involves standard quantitative finance computations. The rolling-window backtest framework adds engineering complexity. Main challenges: obtaining and processing 10-K filings, managing LLM API costs and latency for O(NK) queries per window, ensuring reproducibility of LLM outputs, and handling the CRSP data pipeline. No code is provided.

The paper provides detailed methodology, prompt templates (Figure 3), parameter settings (K=5, 180-day training, 2-month holding, G=5 quintiles, daily rebalancing), and uses DeepSeek-Chat via API with temperature=0 for deterministic outputs. However, CRSP data is proprietary and not freely available. The 768-dimensional embedding model is not explicitly named. No code repository is provided. The LLM API calls introduce non-determinism across model versions, though caching is mentioned. The anonymization of firm names in prompts is well-described.