The core idea
A turnover model estimates a defined departure outcome within a specified future period. It predicts risk, not certainty or the reason a person may leave. Possible inputs require justification: tenure, work patterns and prior experience may be relevant, while absence, pay, location or commute can introduce sensitivity, proxy effects and misleading interpretation. Do not assume that an available field should be used. A useful retention strategy may need only cohort-level analysis and a better process, rather than scores for named employees.
Source and attribution [1]Using it in practice
Define the outcome, prediction date and information available at that date. Use a later holdout period to test generalisation, compare a simple baseline and examine calibration as well as ranking. Check missing data, subgroup errors and changes over time. Decide what supportive, proportionate action could follow a score before building the model. Restrict access and establish a route to question errors. Prediction alone cannot tell you which intervention will change an outcome.
An example, not a reported case
Worked example · illustrative
Suppose a fictional test set contains 1,000 employees and 100 leave within the chosen horizon. A model flags 150 people, of whom 60 leave. Precision is 60/150 = 40%; recall is 60/100 = 60%. Ninety flagged people do not leave, and 40 leavers are missed. Whether that is useful depends on the action, cost and consequences. It would be inappropriate to treat the flagged list as a list of disloyal employees.
What to watch for
Do not use a risk score to deny training, promotion or fair treatment. Historical data can reproduce past inequities, and removing protected fields does not remove all proxies. A model may predict resignation without identifying an effective retention lever. UK data protection and equality duties require specific assessment; this educational guide does not authorise processing or automate employment decisions.
Prevent leakage
Exclude information that only becomes known after the prediction date, such as a recorded resignation or an exit-interview field. Keep related records together when splitting data. A model that looks excellent because it sees the future will disappoint in actual use.
Evaluate beyond accuracy
If only 10% leave, predicting that nobody leaves is already 90% accurate. Examine precision, recall, calibration and the decision threshold. Check performance across relevant groups with enough data and appropriate governance. Report uncertainty rather than only a single favourable score.
Prediction is not treatment effect
The people most likely to leave are not necessarily those most helped by a particular intervention. Test a supportive action with a suitable evaluation design. Consider whether a universal process improvement would be fairer and more useful than individual targeting.
UK governance: check current rules
The ICO confirmed a major DUAA commencement phase on 5 February 2026, while some older guidance remains marked for review. Do not rely on an old Article 22 summary as a complete current assessment. Establish the applicable lawful basis, transparency, minimisation, retention, security and any additional conditions for special-category data. Assess significant automated decisions and meaningful human involvement against the current law and ICO guidance before implementation.
Define the action before the score
Describe what a high predicted risk would lead to and why that action is beneficial and proportionate. A supportive conversation available to everyone differs from withholding development from flagged employees. Consider whether aggregate evidence about schedules, pay or onboarding could support a fairer improvement without individual scoring. If no defensible action exists, stop at the business question rather than building a model because the data is available.
Test a future period honestly
Construct inputs as they would have been known on the prediction date, then evaluate on later outcomes. Check that the same employee's later records have not leaked into training through a careless split. Compare with a simple baseline and report the chosen horizon. A model trained before a major pay or operating-model change may not generalise afterwards. Plan how performance and data quality will be monitored, and what would trigger suspension.
Explain calibration with a practical example
If a model assigns roughly 30% risk to a group of comparable cases, calibration asks whether about three in ten experience the defined outcome over the stated period, across enough observations. It does not mean any named individual is '30% disloyal'. Ranking and calibration answer different questions. Use uncertainty and subgroup checks, and avoid turning a probability into a categorical label in a manager's view of an employee.
Review the whole decision process
Inspect how users interpret the result, whether they can challenge it and whether the proposed support actually helps. Human involvement should be meaningful, with relevant information and authority to change the outcome; a routine approval click is not a quality check. Current UK rules and ICO guidance require a specific assessment before use. The ICO's monitoring guidance is marked for review following DUAA, so treat older summaries cautiously and obtain appropriate advice for the actual processing.
Evaluate the model against a realistic baseline
A model that predicts everyone will stay can appear accurate when departures are uncommon. Examine how many flagged cases actually leave, how many departures are missed and whether predicted probabilities match observed outcomes. Test on later data that were not used to develop the model. Prevent information from after the prediction date entering the inputs. Compare performance with a simpler approach and assess whether the improvement is useful enough to justify the complexity.
Design the response before scoring people
Decide what action a risk estimate would support and whether that action is fair, proportionate and beneficial. Improving workload or onboarding at a group level may be more appropriate than quietly labelling individuals. Do not use a score as certainty, as a reason to withhold development or as a substitute for listening. Review data protection, bias, transparency and governance requirements with current advice. A technically promising model can still be unsuitable if the organisation cannot use its output responsibly.
Member-source update: turn retention analysis into an owned plan
AIHR’s new retention resource links priorities to measures, accountable initiatives and scheduled reviews. Treat that as a planning discipline, not evidence that a named person intends to leave. Original ADM HR exercise: a fictional team reports unpredictable shifts and recurring last-minute changes. Establish a dated baseline for roster changes, voluntary departures and employee feedback. Ask operations to test an earlier roster deadline with a defined budget and review date. Check whether workload and service quality improve alongside retention; also look for burdens displaced onto another team. Keep personal case notes access-restricted and avoid speculative labels such as high flight risk in routine manager dashboards. An individual conversation about aspirations and working conditions is different from covertly predicting resignation. Document what changed, what remained uncertain and the decision to continue, revise or stop.
Take it into your next conversation
Three useful questions.
- Which decision will this analysis change?
- Are the population, denominator, time period and assumptions explicit?
- What alternative explanation or unintended effect must we check?
Related terms
Go to the evidence
Sources & attribution
Further reading: Nadeem Khan and Dave Millner (2023), Introduction to People Analytics: A practical guide to data-driven HR, 2nd edition, PDF pp. 227, 244 and 294–295. Edition & reading notes ↗
[1] NIST: AI Risk Management Framework ↗
[2] ICO: DUAA commencement statement, 5 February 2026 ↗
[3] ICO: automated decision-making guidance consultation and status ↗
[4] Scikit-learn documentation: classification metrics, precision and recall ↗
The core idea is an original summary of the cited work. Application notes, examples and sketches are our interpretations, not quotations or reproductions of the authors’ figures. Publisher records may require access to read the full original work.
Application guidance expanded on 26 September 2026. These practical suggestions and invented examples are editorial explanations; further-reading sources are not the original authors of the historical models unless expressly identified. Original overview review dates are retained. Member-source follow-up on 15 October 2026: the linked resource was reviewed through authenticated access. The application exercise is original ADM HR content; licensed templates and course materials are not reproduced.
Published 2026-09-20 · Reviewed 2026-10-15. Editorial approach