Predictive Analytics in HR

How it works

Every predictive model follows the same pattern. Pick an outcome you care about (left within 12 months, reached proficiency in 60 days, rated a strong performer at first review). Collect past data on people where you know the outcome. Find which inputs were associated with it, such as tenure, commute distance, manager changes, schedule type, pay relative to market or assessment scores. Then apply the model to current people or applicants to estimate the probability of the outcome for each.

The models range from simple (a logistic regression in a spreadsheet tool) to complex (gradient boosted trees or neural networks in an HR analytics platform). Complexity matters less than data quality and a clear decision the prediction will support. It is the forward-looking layer of people analytics.

Common uses

• Attrition prediction. Scoring employees by flight risk so managers can act before resignations. The most common use.
• Predictive hiring. Estimating which applicants are likely to perform well or stay, often from assessments and work history. Tied closely to measuring quality of hire, since you need that outcome data to train anything.
• Workforce forecasting. Projecting headcount needs from demand, attrition and hiring lead times. Contact centers do this daily for call volume and staffing.
• Absence forecasting. Predicting unplanned absence by day, shift and site to improve schedules.
• Offer acceptance and no-show prediction. Estimating which candidates are likely to decline or fail to start, so recruiters can follow up or overbook hiring classes.

Worked example: early attrition in a contact center

A 1,200-agent operation loses about 35 percent of new hires within 90 days. The analytics team builds a model on two years of hires (around 3,000 people) with "left within 90 days" as the outcome.

The strongest predictors turn out to be commute time over 45 minutes, a shift assignment different from the one the candidate said they preferred, a gap of more than 10 days between offer and start, and a low score on the realistic job preview quiz. Interview ratings add almost nothing.

The team does not use the model to reject applicants. Instead it changes the process: shorter offer-to-start gaps, honest shift matching at offer, and extra check-ins in week 2 for new hires the model flags. Ninety-day attrition falls over the next two quarters. That pattern, using predictions to change the process rather than to screen people out, is usually the safest and most useful application.

How to tell if a model is any good

• Validate on data the model has not seen. Test on a later time period, not a random slice of the same one.
• Compare with a simple baseline. If "everyone in their first 90 days is high risk" predicts nearly as well, the model is not adding much.
• Check precision at the threshold you will use. If you will act on the top 10 percent of scores, what share of them actually leave?
• Monitor for drift. A model trained before a pay change, a new site or a labor market shift can decay quickly. Re-check at least every six months.
• Check fairness. Compare predictions and error rates across sex, race and ethnicity, age and disability status where you have lawful data to do so.

Legal and ethical risks

Rules vary by jurisdiction; this is general information.

• Discrimination law. Under Title VII, the ADA and the ADEA, a selection procedure that causes adverse impact on a protected group must be job-related and consistent with business necessity. The Uniform Guidelines on Employee Selection Procedures and the four-fifths rule are the usual starting point for checking selection rates. Proxy variables like zip code or commute distance can carry racial or socioeconomic bias, so using them to screen applicants is risky even if they predict well.
• AI-specific laws. New York City Local Law 144 requires bias audits and notices for automated employment decision tools used in hiring and promotion. Colorado and Illinois have passed laws on AI in employment decisions, and the EU AI Act classes most hiring and workforce management AI as high-risk.
• Privacy. Using personal data for prediction can trigger notice and purpose-limitation requirements under GDPR and state privacy laws such as the CCPA as it applies to employee data.
• Trust. Employees who learn they were scored as a flight risk without being told may become one.

Common mistakes

• Building a model without deciding what anyone will do with the score.
• Training on biased historical decisions, such as past performance ratings, and calling the result objective.
• Reporting accuracy without precision, recall or a baseline.
• Sharing individual risk scores widely with managers instead of using them to target support.