Predictive people analytics can help HR identify teams or cohorts where turnover risk may be increasing before resignations appear in the monthly report. The limitation is not a lack of data. It is that every model can only work with the variables it receives.
A systematic review of 52 peer-reviewed studies found that researchers had used more than 20 machine-learning techniques to predict employee turnover, with most studies relying on supervised learning. A separate 2025 study using 2,932 observations and more than 205 variables tested supervised machine learning on real organizational data to identify determinants of voluntary turnover.
For People Analytics teams, the important question is whether the data stack directly measures the workforce condition leaders care about or infers it from behavior, sentiment, HR events, and other available inputs.
What predictive analytics can already do well?
Turnover models can combine many types of workforce data and identify patterns associated with resignation. The research literature includes models built from compensation, overtime, tenure, role, performance, demographic, and other employment variables.
El systematic review found salary and overtime among the most important factors used across the studies it examined. More recent research has continued testing machine-learning approaches on real organizational datasets, including a 2025 open-access study of employee attrition determinants.
These models can help leaders identify where turnover risk may be concentrated. The limitation appears when leadership wants to understand why a group may be at risk. A model may contain overtime, tenure, absence, performance, or engagement data without directly measuring whether employees feel secure in their jobs, connected to their teams, able to recover from workload, supported by managers, or confident about their future in the organization.
In those cases, the model is working with proxies and related indicators. The underlying wellbeing condition still needs to be measured directly if leaders want to understand how it is changing over time.
That makes the distinction much clearer: the missing variable is the actual human condition, not another operational data point.
Where the blind spot appears
Turnover models are useful for showing where risk may be building based on the data already available. The blind spot appears when the reason behind that risk sits outside the dataset.
If leadership wants to understand whether job security, belonging, recovery, support, or confidence are changing, those conditions need their own consistent measures. Otherwise, the model can flag a pattern without telling leaders what is actually changing in the workforce.
For People Analytics, the practical question is therefore: which workforce conditions are important enough to measure directly alongside the predictive model?
More data does not remove the measurement question
A typical workforce data stack can already tell leaders a great deal.
HRIS data shows tenure, role, movement, and exits. Engagement surveys capture reported employee experience. Attendance systems record absence. Performance systems track goals and ratings. Collaboration data can show changes in work patterns.
Together, those sources create a rich operational view.
The gap appears when leaders need to answer a question that none of those systems measure directly. For example:
- Is workload becoming harder to sustain even though performance has not fallen?
- Do employees feel less secure after a restructuring even though voluntary turnover is still low?
- Is belonging weakening in one cohort while the company-wide engagement score remains stable?
- Is confidence or perceived support improving after an intervention?
- Are broader wellbeing conditions moving before absence, retention, or performance changes become visible?
Those are different questions from “Who is likely to leave?”
Turnover prediction and wellbeing measurement serve different decisions
A turnover-risk model helps answer where departure risk may be increasing based on the variables available to the model.
A longitudinal wellbeing measure helps answer how specific aspects of workforce wellbeing are changing over time.
The two can be used together.
Equipos de análisis de personas can compare changes in wellbeing with absence, engagement, retention, performance, or other business measures. The value is not in claiming that one wellbeing score predicts resignation. The value is in adding a direct measure of a workforce condition that might otherwise be inferred indirectly.
That gives analysts a clearer way to separate three things:
- what the organization directly measured
- what a model inferred from available variables
- what business outcome eventually occurred
Keeping those categories separate makes the analysis easier to explain and reduces the risk of overstating what a predictive score means.
The real opportunity is better variable design
Many organizations already have segmentation, dashboards, predictive models, and continuous reporting. Adding another dashboard does not solve the underlying problem.
The stronger question is whether the analytics model contains the right variables for the decisions leaders are trying to make.
If the organization wants to understand workforce wellbeing, it needs a consistent way to measure wellbeing itself, rather than relying only on proxies such as absence, turnover, engagement, or work patterns.
If the organization wants to understand post-layoff recovery, it needs direct measures of conditions such as security, support, workload strain, belonging, and confidence.
If the organization wants to understand whether a wellbeing investment is associated with change, it needs repeated outcome measures that can be compared over time and across relevant cohorts.
The quality of the insight depends on the quality of the variables.
A practical review for People Analytics teams
When reviewing a workforce analytics stack, map each important business question to the data that answers it.
For each question, ask:
- Are we measuring this condition directly or using a proxy?
- If it is a proxy, what assumptions are we making?
- Which groups can we compare?
- Can we see movement over time?
- What business outcome would we expect to examine alongside it?
- What would the data still not prove?
That exercise often reveals where the stack is genuinely strong and where leaders are asking one dataset to answer a question it was never designed to answer.
The goal is not simply more prediction. It is a clearer connection between the question, the variable, the evidence, and the decision.
