Predictive attrition modelling uses past HR data to estimate each employee’s chance of resigning in a coming period, usually shown as a flight-risk score or band. It is the most common use of people analytics, and the hard part is not building the model but deciding what you do when someone is flagged.

What goes into a flight-risk model

The model learns from employees who stayed and those who left, so it can only use what your HRMS already holds. Typical inputs:

• Time since last promotion or increment, and pay position in the band, often as a compa-ratio.
• Recent manager change, transfer or team reorganisation.
• Tenure, especially the stretch around one year and around ESOP vesting dates.
• Rating trend, overtime hours and long commute, where those are recorded.
• Exits among the person’s immediate team in the last six months.

Model types range from plain logistic regression, which gives a probability and shows which factors pushed it up, to tree-based methods such as random forests, which often predict a little better but are harder to explain. Survival models estimate when someone might leave, not only whether. Train only on voluntary exits; layoffs and terminations are a different story, as the voluntary and involuntary split explains.

False positives, with numbers

An illustrative case: 1,000 engineers, of whom 150 will resign in the next year. The model flags 200 people as high risk and 100 of them do leave. It caught 100 of 150 leavers (67%), but half the flagged list, 100 people, were never going anywhere. Fifty leavers were missed entirely. That is a decent model, and it still means a “high risk” label is a coin toss for any one person.

The ethics of acting on a score

Use scores to direct attention, never to decide. A flag should trigger a manager conversation or a stay interview, not a withheld promotion or an early exit. Show managers bands for their team, not raw probabilities, and keep the score out of appraisal and pay discussions. Check whether the model flags women returning from maternity leave or one location far more often, since that can encode past bias. India has no AI-specific employment law as of October 2026, and MeitY’s November 2025 AI governance guidelines are not binding, but section 8(3) of the DPDP Act will require complete, accurate and consistent data where it is used for a decision affecting the person, once its duties start applying on 13 May 2027.

Frequently asked questions

Do we need a data scientist?

Not to start: some HRMS products include a flight-risk score; you need someone who can question it and check it against actual exits.

How much history is enough?

Two to three years of clean exit data with reasons, ideally a few hundred voluntary leavers; with only 30 exits, the model mostly learns noise.

Should employees be told about the model?

Yes, describe it in your privacy notice and explain how scores are used; a secret score that leaks does more damage to trust than the attrition it was meant to prevent.

Put a rupee figure on each prevented exit with the cost of turnover calculator before deciding how much to spend on retention.