Introduction: Large factor libraries in the realm of quantitative investment run into obstacles like information leakage, redundancy, and multiple testing bias and even the reproducibility as well as the implementation of strategies pose difficulties. Methods: This paper proposes an automated system to generate factors using type-safe operators syntax. Nonlinear screening with High-Sensitive Inference (HSIC-Lasso) and redundancy suppression were used. The stability analysis was performed by using Purged/Embargo nested cross-validation and Deflated Sharpe Ratio (DSR). The management of both decentralized execution and audits was done on the Arrow/Parquet and Ray/Dask systems. Results: The system based on the WRDS CRSP /Compustat combined database (2014-2024) yielded an average RankIC of about 0. 031 and a combined IR of about 1. 04 as well as an average DSR p-value of less than 0. 01. The stability of the IC indicator distribution was observed across various market cycles and trading activity was greatly reduced. Regardless of whether 10⁵ candidates were added, the growth of the computational scale preserved linearity. Conclusion: Our work has provided us with an opportunity to combine statistical strength and engineering scale to develop a reproducible path to mine factors. Our plans are to enhance the quality of adaptive evaluation, causal screening, and multi-market migration.
Jiahe Sun (Thu,) studied this question.