Discovery under Hypothesis Redundancy: A Geometric Theory of Discovery Bottlenecks

Battle Count: 69

Highly relevant to quantitative trading. The paper uses A-share factor discovery (5,647 stocks, 2010-2026) as a primary testbed, with ICIR, IC, Sharpe ratios, transaction cost analysis (10-50 bps), risk neutralization (industry, size, joint), and walk-forward validation. The framework directly addresses the factor zoo problem and provides a diagnostic for when LLM-guided alpha discovery adds value over structured factor enumeration. The top discovered factors (gapupratio_20d, mom_vol, ixret5xret10, ret5plus_vol, volratio) are actionable trading signals. The budget-allocation diagnostic has direct implications for quant research resource allocation.

Moderate to high complexity. Requires: (1) LLM integration for non-local seed generation with factor taxonomy and knowledge base, (2) structured enumeration engine with multiple operators (product, ratio, difference, rolling correlation, percentile rank, z-score), (3) GPU-accelerated evaluation sandbox with timeout/memory management, (4) spectral analysis of correlation matrices (effective rank computation), (5) ICIR-based verification with redundancy filtering, (6) market-cap segmentation, (7) temporal train/validate/test splits, (8) risk neutralization and multiple testing corrections (Romano-Wolf bootstrap). The theoretical framework itself is mathematically sophisticated (spectral theory, leverage bounds, concentration inequalities).

The paper describes data sources (Tushare Pro API), splits, metrics, thresholds, Monte Carlo counts, and protocols in detail. Implementation details are in Appendix A. Scripts, cached tables, and manifests are stated to be released upon acceptance. However, raw A-share market data is commercial and non-redistributable. The LLM used (deepseek-v4) and specific prompting strategies are described but full reproducibility depends on data access. No GitHub repository URL is provided in the extract.