A Comparative Study of Machine Learning Regression Models for Predicting Remote Worker Productivity: Accuracy, Robustness, Efficiency, and Interpretability

Abstract

Remote and hybrid work has become the norm, yet the tools organizations use to gauge employee productivity have not kept pace, leaving many to lean on indirect proxies of uneven accuracy. This study investigates whether present-day Machine Learning (ML) regression can reliably measure and forecast remote-worker productivity, mapping out the trade-offs that should guide model choice in practice. Nine regression algorithms-linear, tree, ensemble, kernel, distancebased, and neural-are benchmarked on a public dataset of 1,500 remote employees described by 27 predictors, all passed through one shared preprocessing pipeline and scored on four axes: accuracy (MAE, RMSE, $\mathbf{R}^{\mathbf{2}})$, robustness (5-fold crossvalidation and per-quintile stratified RMSE), efficiency (wallclock training time), and interpretability (native importances and SHAP values). Gradient Boosting achieves the top heldout $\mathbf{R}^{2}$ of $0.9257($ RMSE $=3.90)$, MAE $\left.=3.07\right)$ alongside a crossvalidation $\mathbf{R}^{2}$ of $\mathbf{0. 9 3 1 0} \pm \mathbf{0. 0 0 9 1}$, while Lasso Regression is statistically indistinguishable from it $(\mathrm{R}^{2}=0.9250)$ despite training about fifty times faster. Three output-quality features-Efficiency Rating, Quality Score, Task Completion Rate-supply 96.2% of the leading model's importance mass. Withholding these features in a leakage-aware re-evaluation drops held-out $\mathbf{R}^{\mathbf{2}}$ by $\mathbf{6 7. 7} \%(\mathbf{0. 9 2 6} \rightarrow \mathbf{0. 2 9 9})$; a Friedman-Nemenyi analysis across 30 repeated-CV folds confirms the differences are significant $\left(p<10^{-35}\right)$, and a budgeted grid search shows the ranking is not an artefact of unfair defaults. These results offer organizations concrete guidance for deploying ML pipelines in productivity analytics.