金融研究arXiv q-fin7 Oct, 12:00 pm SGTSIGNAL CDC27D

FactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining

FactorBench is a portfolio-aware benchmark that compares roughly five thousand mined factors from nine automated factor mining methods across five equity markets. Its novelty lies in a shared data and evaluation contract that connects heterogeneous discovery algorithms to common signal combination and portfolio construction procedures and traces outputs across factor validity, temporal generalization, predictiveness beyond risk and style exposures, and pool distinctness, help

01 ABSTRACT

The paper introduces FactorBench, a portfolio-aware benchmark comparing about five thousand factors mined by nine automated methods across five equity markets. It traces outputs across factor validity, temporal generalization, predictiveness beyond measured risk and style exposures, within- and across-method pool distinctness, composite-signal quality, and after-cost long-only and long-short portfolio performance. The authors report that no paradigm consistently dominates; this is their conclusion based on the benchmark.

02 KEY FINDINGS

Covers roughly five thousand mined factors, nine automated methods, and five equity markets under a shared data and evaluation contract.
Evaluates factor validity, temporal generalization, predictiveness beyond risk and style exposures, method-pool distinctness, and after-cost portfolio performance.
Concludes that no paradigm consistently dominates across all evaluated dimensions.

03 RELEVANCE

Many automated factor mining approaches exist, but their cross-paradigm generalizability, distinctness, and economic usefulness have lacked a unified comparison. FactorBench offers a portfolio-aware evaluation framework to help researchers judge whether methodological advances translate into signal quality and after-cost portfolio performance, without constituting investment advice.

04 EVIDENCE BOUNDARY

The benchmark covers nine automated mining methods, five equity markets, and about five thousand factors, so conclusions depend on the selected methods, markets, evaluation contract, factor-pool construction, and backtest settings. The reported absence of a consistently dominant paradigm is not shown to be robust across all markets, periods, and cost assumptions; the evidence is limited to the evaluated sample and benchmark design and should not be generalized to all factor mining methods or markets.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG · ARXIV Q-FIN

Read original ↗