What drives risk
Feature importance (random forest)
Tenure and engagement were the strongest predictors. Early-career employees and anyone whose engagement is sliding are where intervention pays off most.
How risk is spread across the workforce
Explore predicted risk
Predicted attrition probability by age
Each dot is one employee in the synthetic sample. Pick a department to focus on it.
Employees with frequent overtime and low job satisfaction show the highest predicted risk in this sample.
Logistic regression coefficients (second model)
Coefficients are not standardized, so Monthly Income looks large only because it is measured in dollars. Read this chart for direction of effect, not size; the random forest importances above are the comparable view.
Data and method
- Data: anonymized HRIS fields including tenure, performance ratings, compensation bands, engagement survey scores, promotion history and absence patterns. Personal identifiers removed.
- Features: cleaned missing values, normalized numeric fields, encoded categories and engineered signals such as time since last promotion.
- Model: compared logistic regression, random forests and gradient boosting; chose a random forest for its balance of accuracy and interpretability. Final F1 0.72, AUC 0.81 on a held-out test set.
Ethics
Sensitive attributes (age, gender, ethnicity) were excluded from the production model and it was audited for disparate impact. Risk scores fed a broader retention conversation, never an automatic decision, and the work was reviewed with legal and privacy partners.
Results and lessons
HR teams aimed stay interviews and development opportunities at people flagged as high risk. Within six months, voluntary attrition was down 28% against the prior year, and the model surfaced systemic issues, such as a lack of mentoring, that shaped broader programs.
- Data quality matters. A reliable pipeline mattered as much as the model.
- Collaboration is the job. HR partners, data engineers and leaders had to trust the output.
- Models drift. Retraining and monitoring were scheduled from day one.
About the data. The exploration chart uses the synthetic employee sample published with the original interactive page (age is shown there to explore the sample; it was excluded from the production model). Importance and risk-distribution figures come from the original interactive dashboard. The 28%, 0.72 and 0.81 outcomes are carried over unchanged from the original case study.
Rebuilt from:
- predictive_attrition_case_study.html (write-up, outcomes)
- interactive_dashboards/ai_attrition_dashboard.html (feature importance, risk distribution)
- interactive_cases_refined/attrition_case.html (employee-level sample, coefficients)
More reports
Reducing 90-Day Attrition
Predictive and prescriptive analytics to decide where a limited onboarding budget does the most good for early retention.
Open reportWorkforce Planning & Forecast
Twelve-month headcount scenarios by department under different attrition and hiring rates, plus a live Tableau forecast.
Open reportMeasuring Engagement Impact
Does engagement move productivity and turnover? A department-by-department look at the data, with the honest answer.
Open report