Workforce & Retention

Flight Risk Score

Also known as Attrition Risk Score, Turnover Risk Model

A flight risk score is a modelled probability that a given employee will voluntarily leave within a defined horizon, usually the next six or twelve months. Unlike turnover rate, which reports what already happened, it is a forward-looking estimate built from tenure, engagement, compensation position, manager history and internal movement data.

Formula

Typically a logistic regression or gradient-boosted model: P(leave within horizon) = f(tenure, time in role, compa-ratio, engagement trend, manager change, promotion recency, commute, peer attrition)

There is no universal formula — the value is in a model fitted to your own historical exits. A model trained on someone else's data will not transfer, because the drivers differ by organisation.

Worked example

Scenario
A model trained on three years of exits scores a 200-person engineering group, flagging the top decile as elevated risk.
Calculation
20 employees scored above the 0.35 probability threshold; of those, historical precision suggests roughly half would leave within 12 months absent intervention
Result
A shortlist of 20 for manager conversations, rather than a 200-person retention campaign

What good looks like

Judge the model on precision within the top risk decile rather than on overall accuracy. Because base attrition rates are low, a model that predicts 'nobody leaves' will score highly on accuracy and be worthless. Useful models typically identify a top decile containing several times the base rate of actual departures.

Why flight risk score matters

Retention budgets and manager attention are finite. A ranked risk list directs both at the people most likely to leave and most costly to lose, which is a materially better use of effort than a company-wide intervention. It converts retention from a reactive exercise into a planned one.

How to improve it

  • 1Train on your own exit history — at least two years, ideally three — rather than importing assumptions about drivers.
  • 2Weight the output by role criticality and replacement cost so the list reflects impact, not just probability.
  • 3Give managers the drivers behind each score, not just the score. A number with no explanation produces no action.
  • 4Re-fit periodically. Drivers shift with the labour market, and a model trained on a hiring boom misreads a downturn.
  • 5Measure whether flagged employees who received intervention stayed at a higher rate than flagged employees who did not.

Common mistakes

  • Using accuracy as the evaluation metric on an imbalanced outcome, which rewards a model that predicts no attrition at all.
  • Including protected characteristics as features, directly or via close proxies — a legal and ethical problem as much as a modelling one.
  • Exposing individual scores without governance, which damages trust far faster than the model creates value.
  • Treating the score as a verdict rather than a prompt for a conversation.

Frequently asked questions

Is it ethical to score employees on flight risk?

It can be, with governance. The defensible version uses the score to trigger a supportive conversation and excludes protected characteristics and their proxies. The indefensible version uses it in promotion, pay or assignment decisions, which converts a retention tool into discrimination.

How much data do I need to build a flight risk model?

Realistically two to three years of exit history and a few hundred separations. Below that, a well-designed rule-based risk flag — tenure band, time since last promotion, compa-ratio, engagement trend — will outperform an under-trained model.

Related metrics

Go deeper

See flight risk score in your own data

Connect your HR system and PeoplePilot builds this metric — and the rest of the library — segmented and updated continuously.