Dual-target modeling of team performance using regularized regression with expanding-window rolling validation and engineered feature ablation: methodological insights for small-sample digital public health analytics

Accurate estimation of health outcomes from small-sample time-series data remains a persistent challenge in digital public health, particularly when data are limited and tailored analytical frameworks are scarce. This study addresses this gap by proposing a systematic machine learning pipeline that, while demonstrated on Women's National Basketball Association (WNBA) team performance data, may offer a useful methodological reference for small-sample public health estimation tasks. We collect team-level statistics from 20 WNBA seasons (2006–2025) and construct nine engineered indicators capturing quarter-level scoring dynamics, positional synergy, and defensive efficiency. Two target variables are modeled separately: Points Per Game (PPG) as a process indicator and Win Percentage (WIN%) as an outcome indicator. Eight regression algorithms are evaluated under an expanding-window rolling validation framework that respects temporal ordering. Results demonstrate that regularized linear models consistently outperform ensemble tree-based and kernel methods for both targets: Lasso Regression achieves R 2 = 0.9443 for PPG, while ElasticNet Regression achieves R 2 = 0.9287 for WIN%. Ablation experiments reveal that only three of nine engineered features contribute positively to PPG modeling; notably, an optimized 12-feature subset outperforms the full 18-feature set ( R 2 = 0.9497 vs. 0.9443), demonstrating that feature selection is as important as feature creation. For WIN%, the advanced metrics in Set B already capture most variance, with engineered features providing minimal marginal gain (Δ R 2 < 0.001), confirming that net rating overwhelmingly determines win percentage. Beyond sports analytics, the proposed expanding-window validation framework and systematic feature ablation approach may provide a rigorous template for digital public health research facing similar small-sample constraints, such as infectious disease surveillance, community health intervention evaluation, and population-level behavioral modeling, where temporal integrity and feature parsimony are critical. The methodological transferability of the framework is further validated through an external experiment on under-five mortality rate estimation using cross-country panel data, where Gradient Boosting achieved R 2 = 0.922 with progressive feature engineering.