What an HR record can and cannot say about who leaves.

HR teams are now shown attrition analytics with a steep curve, a high accuracy figure and a list of names. What the slide rarely says is what the figure was measured against, who the analysis misses, and what a century of research already knew about the limits of the record.

What a hundred years of research says

Hom, Lee, Shaw and Hausknecht's 2017 review traced a century of turnover theory and noted the field's recent turn towards the psychology of staying rather than leaving. The largest quantitative summary is Rubenstein and colleagues' 2018 meta-analysis: 316 studies, 57 antecedents, about 1,800 effect sizes, with an average base rate of 23.5% and an average gap of roughly sixteen months between when a predictor was measured and when people quit.

The ranking is the point. Corrected correlations with actual quitting were strongest for intent to quit (.56), job search (.40), organisational commitment (−.29), job satisfaction (−.28) and job embeddedness (−.26). Age (−.21), tenure (−.20) and pay (−.17) were weaker. The variables that matter most are the ones an HR system never holds, and the ones it holds matter less. A 2022 review of 1,375 turnover studies added a caution: much of the field measures the intention to leave as a proxy, not leaving itself.

What the record can do

That is not a reason to ignore the record. A 2025 study in a large Danish municipality, combining administrative data with surveys, found that demographic variables carried the most weight for actual turnover and that work-environment and attitude measures did not significantly improve accuracy. A 2025 analysis of 121,883 financial professionals over seventeen years found that people were 23% more likely to leave when more than 30% of their peers had departed within six months, a signal that is entirely in the record.

That sets the bar for any richer analysis. Add moves, pay against peers, pay growth, ratings, manager changes and resignations around each person, and the ranking rarely improves much on tenure and location alone. What the richer factors buy is explanation: why this person, and what might help, which is what an HR business partner needs for a conversation. Published research on each factor's direction and rough size is a sounder starting point than anything learned from a few months of local data.

Test out of time, or the test means nothing

Kapoor and Narayanan's 2023 survey found leakage in 294 papers across 17 fields, and named temporal leakage as the most common kind: a test set containing data from before the training set, so the analysis is built on data from the future. In one case they re-ran, fixing the leak cut the reported advantage of machine learning over a simple regression from 0.14 to 0.01 in AUC. A 2024 BMJ methods paper is blunter: randomly splitting a dataset “is often erroneously perceived as a methodological strength, it is not”. A February 2026 study measured the same effect on the same data: 0.775 under a random split, 0.730 under temporal validation.

The only honest test trains on earlier periods and scores a later one, month by month. Freeze the inputs several months before each outcome, shuffle the outcomes as a control, and publish who the analysis misses. Expect them to be experienced people leaving for reasons the record never held: an outside offer, a partner's move, a conversation that never reached a field. No analysis of HR data will see those people coming, and saying so belongs in the results.

The law has caught up with scoring

In SCHUFA (C-634/21, December 2023), the EU Court of Justice held that a probability value is itself an automated decision under Article 22 of the GDPR when the party using it relies on it strongly. In Dun and Bradstreet Austria (C-203/22, February 2025) it added that a person is entitled to an explanation of “the procedure and principles actually applied” to their data, and that handing over an algorithm does not count. The EU AI Act lists systems that evaluate workers as high-risk, and under Article 6(3) any such system that profiles people is always high-risk. In Germany, a system able to monitor behaviour or performance also needs the works council's agreement.

The EDPS said in September 2025 that meaningful human oversight is “active involvement that improves the quality of the decisions”, not “a merely procedural formality”, and named recruitment as a field where automation bias turns a decision-support tool into a de facto automated decision-maker. A September 2026 survey of 22,573 US employees found that only one in five trusts their leaders a great deal to use workforce data responsibly.

Design the output for the conversation

1. Levels, not percentages. A 23% probability invites false precision; a level says who to talk to first.
2. Reasons that add up, in plain language, shown as hypotheses to check rather than verdicts.
3. Supportive actions only. A development conversation, a pay review, onboarding help. Nothing is refused or removed because of a level.
4. Names stay with HR business partners. Managers see teams above a minimum size.
5. Fairness as a release gate: if errors fall unevenly across groups, nothing goes live until it is fixed.
6. A monthly check of expected against actual, so the analysis earns its place or loses it.

What to ask before you buy one

1. Was it tested out of time, on a period it never saw?
2. How does it compare with a baseline that knows only tenure and location?
3. Who does it miss, and what happens when a fairness check fails?
4. What does a manager see, and can an employee ask for the reasons that apply to them?