Abstract

In this research, the dual prediction and prescription model is developed and validated so that this model can not only predict employee turnover risk, but it also proposes the appropriate retention interventions, which apply across industries. On the basis of the IBM HR Analytics Attrition dataset (n=1,470), we preprocessed demography, job and satisfaction variables and trained three machine-learning classifiers, Random Forest, Logistic Regression and XGBoost to predict voluntary turnover. XGBoost model recorded the best discrimination (AUC=0.87), sensitivity (0.76), and specificity (0.81), which signifies strong predictive power. Analysis of feature-importance was conclusive with time, rate of frequent business travel and compensation measures as major causal factors of attrition, but the factors had different levels of prominence per sector, which required a customized approach to intervention. In order to operationalize predictions, we paired the best predictors with evidence-based HR lever and came up with a dynamic monitoring solution that included real-time risk scores and an employee engagement survey to operationalize the prediction. The dataset exhibited class imbalance, with attrition representing a minority proportion of the sample. To address this, model development incorporated ROSE resampling and threshold tuning to improve sensitivity to true leavers, achieving strong discrimination (AUC = 0.95) while reducing false negatives. We have prescriptive recommendations which are specific to different job functions based on how to cope with the workload demands of front office jobs, as well as how to streamline promotion channels within technical divisions, with an A/B testing strategy which would help to improve the effectiveness of applications. The study contributes to the understanding of usingworkforce analytics with a clear example on how predictive analytics lead to implementing actionable HR policies so that attrition becomes not a reactive practice in talent stewardship, but an active one. Future researchers are advised to use longitudinal and post-pandemic data to trace the changing work pattern and include mixed-method implications to draw quantitative results into detail.

Publication Date

12-2025

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Sanjay Modak

Advisor/Committee Member

Ayman Ibrahim

Campus

RIT Dubai

Download

Share

COinS