MKS Publication

Chapter 3

Predicting Employee Turnover with Explainable AI: A Manufacturing Case Study

Abstract/Preface

Voluntary turnover is costly for labour-intensive manufacturing in emerging economies, and machine learning is increasingly proposed as a way to anticipate which workers are most at risk. A real, end-to-end explainable machine-learning study of voluntary resignation in a large textile manufacturer in Thailand draws on 1,160 employee records, of which 232 are voluntary leavers. Gradient-boosted decision trees are combined with Shapley Additive explanations (SHAP) to predict and explain resignation, with interpretable baselines for comparison and a group check across gender. The model is genuinely predictive (cross-validated ROC-AUC of 0.78, compared with 0.62 for logistic regression), but its most useful contribution lies in the explanations it reveals. Contrary to the common assumption that low pay drives attrition, the workers who left earned substantially more than those who stayed, and higher pay raised predicted resignation risk; the conventional stress indicators of overtime, absenteeism, and age did not distinguish leavers from stayers at all. The decisive evidence is behavioural: 183 of the 232 leavers (79%) subsequently returned to the firm at their former wage, reporting that the outside offers that drew them away were either too distant or simply untrue. The study’s central lesson is that, in this workforce, resignation is driven by external labour-market pull and misinformation that internal HR data capture only indirectly, so the practical value of an explainable model lies less in the score than in redirecting retention effort toward marketable workers in exposed departments and toward countering false external promises with transparent information about total compensation.