Key Figures
What Machine Learning Actually Does
Machine learning, at its core, is pattern recognition at scale. A turnover prediction model examines the workforce data of thousands of employees over time, learning which combinations of factors preceded past departures. It then applies these learned patterns to current employees to assess their individual turnover risk.
This is fundamentally different from simple rules like "employees with more than 10 sick days are likely to leave." ML models consider hundreds of variables simultaneously, identifying complex interactions between factors that no human analyst could detect across a large workforce.
The Signals That Matter
Turnover prediction models typically consider several categories of signals. Engagement indicators include changes in attendance patterns, overtime behaviour, and participation in voluntary activities. Workload signals include hours worked, consecutive shift patterns, and rest period adequacy.
Tenure and career factors include time since last promotion, length of service, and role changes. External factors include industry turnover rates, competitor activity, and regional labour market conditions. The model learns which combinations of these signals are most predictive for each specific workforce.
Understanding Accuracy
When a model reports 85% accuracy, it means that 85% of the time, its prediction aligns with what actually happens. In turnover prediction, this typically means that the majority of employees identified as high-risk do go on to leave within the predicted timeframe, and the majority of employees identified as low-risk remain.
No model is perfect, and false positives (employees flagged as at-risk who stay) and false negatives (employees who leave without being flagged) both occur. In practice, a false positive — having a retention conversation with an employee who was not actually going to leave — is relatively low-cost and may even be beneficial. A false negative is more costly, but the overall reduction in unexpected departures still delivers significant value.
Ethical Considerations
Predictive workforce analytics raises important ethical questions. How should turnover risk information be used? Who should have access? How do we ensure the model does not encode biases?
Responsible implementation requires clear governance. Predictions should be used to support employees, not to discriminate against them. Risk scores should never be used as the sole basis for employment decisions. Transparency about how predictions are generated builds trust with both managers and employees. The goal is to create better workplaces, not surveillance systems.