A Working Guide to Attrition Modeling for HR Leaders

Most attrition models built inside mid-market HR functions fail in the same way. They predict the wrong unit (company-wide rates instead of team-level risk), at the wrong time horizon (a yearly figure when the decisions are quarterly), with features that are correlated with exit but not useful for intervention.

This guide is opinionated about what to build instead. It is based on what we have seen work across 47 customer deployments. Your situation will differ on the margins, but the broad shape of the model is reliably the same.

Pick the right unit and horizon first

The single biggest lift in attrition modeling comes from picking the right thing to predict. Company-wide annual attrition is a vanity metric. It tells you nothing actionable. What you actually want to predict, in most cases, is:

• The probability of voluntary exit, at the individual employee level, over the next 90 days.
• Aggregated to the team level (5 to 25 employees) for HR business partner action.
• Refreshed weekly, with stable predictions that do not flip on small data changes.

This unit and horizon match the decision cycle of an HR business partner. It is too granular to use for board reporting (use the rolled-up trailing figure for that), but it is the level at which an HR leader can actually do something about it.

The four features that do most of the work

From thousands of candidate features we have tested across customer deployments, four consistently account for the majority of the predictive signal on 90-day voluntary attrition. They are, in rough order of importance:

1. Tenure-bucketed exit rate, at the role-and-department level

The base rate of exit for "Software Engineers in their second year of tenure inside the Platform Engineering department at this specific company" is almost always the largest single feature in a working attrition model. Without this base rate, you will be over-predicting risk for stable cohorts and under-predicting it for high-churn ones.

2. Compa-ratio relative to current market

The ratio of an employee's salary to the current external market rate for their role, geography, and experience band. A compa-ratio below 0.90 produces a strong exit signal in our data. Above 1.0, the signal goes flat — paying people 10 percent above market does not meaningfully reduce risk past the threshold.

3. Manager change in the last 12 months

Employees who have had a manager change in the last 12 months — particularly those whose manager left voluntarily — are 1.7 times more likely to exit in the next 90 days, controlling for other features. This effect is well-documented in academic literature and consistently replicates in our customer data.

4. Engagement survey signal (specifically: the four questions discussed in our engagement survey research)

Engagement signals contribute meaningful predictive lift when they are available, but they are usually only refreshed quarterly. We weight them inversely with survey recency — a survey from three weeks ago counts for more than one from three months ago.

Features we have tested and dropped

For completeness, a non-exhaustive list of features we have tested and found do not add meaningful predictive value once the four above are included:

• Performance review scores. Largely subsumed by the base rate feature.
• Time since last promotion. Useful in isolation, but redundant with tenure-bucketed exit rate.
• Slack or Teams activity. Surface-level proxies that mostly capture role behavior, not exit intent.
• Calendar density. Same problem.
• Vacation balance. Counter-intuitively, a weak signal — high balances correlate with both burnout and stable tenure.

Model class

For most mid-market deployments, a gradient-boosted decision tree (XGBoost or LightGBM) outperforms more sophisticated approaches once you account for interpretability, stability across refreshes, and the realities of running on the data volumes a 2,000-employee company actually has.

Neural network approaches we have evaluated produced 2 to 4 percent AUC improvements in our largest deployments. In smaller ones, they underperformed simpler models due to data sparsity. The interpretability cost is significant, and HR partners need to be able to explain why the model is flagging a specific employee.

What to do with the predictions

The hardest part of attrition modeling is not the model. It is what happens after the predictions land in someone's inbox. Three things we have seen work, and one to avoid:

1. Route flags to the HR business partner, not the manager. Direct delivery to managers produces awkward, sometimes counter-productive conversations. HRBPs can triage and decide what action is appropriate.
2. Pair the prediction with the top three driving features. "This person is high risk because their compa-ratio is 0.84 and they have had a manager change" is actionable. "This person is high risk" is not.
3. Track interventions and outcomes. Without an intervention tracking layer, you cannot learn whether your follow-up actions work. Most teams skip this and pay for it later.
4. Avoid: building a "retention dashboard" with no escalation path. A dashboard nobody looks at is the most common failure mode.