Features for Screening Versus Features for Clustering: A Methodological Distinction in Systematic Equity Investment

Abstract

Systematic equity strategies typically apply the same feature space to both stock selection and portfolio clustering. This paper establishes that screening and clustering address different questions and require different features. On a universe of 280 US equities above $30 billion market capitalisation, a Tree-structured Parzen Estimator identifies 76 stocks through 17 return-based features; adding 12 Compustat financial ratios does not improve selection (paired t-test p = 0.117; TOST equivalence at ±3% annualised p TOST = 0.044). By contrast, 14 recovery-mechanism features combining business structure with price-derived characteristics enable cluster-level mean-variance to outperform equal weighting (p = 0.004), while 379 characteristics from Jensen et al. (2023) produce only outlier detection within the screened portfolio. As external validation, the methodology has been deployed with live capital since February 2023, producing 38 months of realised returns consistent with the backtested properties. The finding is that appropriate feature spaces depend on the question being asked, and that pipeline-stage feature differentiation is empirically meaningful under both historical test and live deployment.

Keywords: stock screening, portfolio clustering, feature selection, Tree-structured Parzen Estimator, recovery pathways, cross-sectional characteristics, permutation importance, mean-variance optimisation, hierarchical clustering JEL Classification: G11, G12, G14, G17, C55, C58 Features for Screening vs Features for Clustering