Can employee turnover really be predicted?

Yes, employee turnover can be predicted better than chance inside a defined dataset. That is the modest conclusion the evidence supports. It does not follow that a model will travel well to another employer, identify a cause or tell a leader which intervention will work.

Performance depends on the problem you actually measure

Ribes, Touahri and Perthame studied about 1,000 employees in one critical job category over two years. Voluntary turnover was near 20 percent, so the authors treated class imbalance as a central modeling problem. A random forest with ROSE resampling reached ROC-AUC 0.95 in their held-out comparison, while an uncorrected LDA model reached 0.73.

That is a strong within-case result, not a universal attrition score. The role, company and available variables were all specific. The study is valuable because it makes the context visible and because it tests a path from ranking risk to designing targeted policies.

More complex did not mean more useful

Derrazi and Pourmostafa Roshan Sharami compared XGBoost, LightGBM, SAINT and SAINT-tree hybrids on one public dataset with 74,498 records. XGBoost and LightGBM produced ROC-AUC values of 0.8529 and 0.8521. Both outperformed the tested transformer and hybrid variants.

The result matters beyond the leaderboard. The standalone tree ensembles were also faster and easier to interpret at the feature level in this experiment. A more elaborate architecture can be justified when it produces a material gain. Here it added cost without buying better discrimination or clearer explanations.

Accuracy can hide the failure you care about

Mohiuddin and colleagues used IBM's synthetic dataset of 1,470 employee records, only 237 of which represented attrition. Their XGBoost model reported 89.12 percent accuracy, but the attrition problem remained difficult: the reported F1 score was near 40 percent in the weighted-feature setup. The paper itself warns that accuracy alone can mislead when classes are imbalanced.

For an employer, the practical evaluation needs to start with the decision. False positives can lead to intrusive attention or wasted retention offers. False negatives leave preventable risk unseen. Precision, recall, calibration and performance by employee group are more informative than one headline accuracy number.

An explanation is not an intervention

SHAP values, what-if tools and counterfactual explanations can make a model inspectable. They can show which variables were important to a prediction or which changes would cross its decision boundary. They cannot establish that changing salary, satisfaction or tenure will cause a person to stay.

That distinction is especially important in the studies by Mohiuddin et al. and Artelt and Gregoriades. Both use synthetic IBM data. Their scenario and counterfactual outputs are demonstrations of model behavior, not evidence-backed policy effect sizes.

From model score to responsible use

A defensible turnover model has a narrow stated purpose, a documented population, time-aware validation and a baseline simple enough to challenge. It is monitored for calibration and group-level error. Most importantly, it supports a humane conversation about working conditions rather than automating a judgment about an employee.

The RetainScore diagnostic follows the same boundary. It is a structured risk conversation about workplace factors a leader can examine. It is not an individual flight-risk classifier and does not claim to predict who will resign.