Predictive HR analytics usually means one of two things: forecasting an aggregate, such as how many people a division will lose next year, or scoring individuals for a risk such as leaving. The first is ordinary forecasting and is useful. The second is a decision about a person made partly by a model, and it carries obligations.
Aggregate forecasting is the safe and useful half
Turnover by team, hiring demand against a plan, absence seasonality. These support resourcing decisions, need no judgement about any individual, and can be explained to anybody who asks. They also need considerably less data than vendors suggest, because the useful signal is usually in a few years of clean history rather than in exotic sources.
Scoring individuals changes the question
A flight risk score attached to a named person will influence how that person is treated, whatever the policy says. That makes it a decision aid about an individual, with duties around transparency, fairness and the ability to explain the basis. If the vendor cannot describe how a score is produced, you cannot defend a decision that was influenced by it.
Check the model does not encode the past
A model trained on who was promoted historically will reproduce whatever biases existed in those decisions, and it will do so consistently and at scale. Monitor outputs by group, require the vendor to explain the features used, and be prepared to switch it off. This is ordinary diligence rather than a reason to avoid the field.
Questions people ask about predictive hr analytics
How much data is needed?
For aggregate forecasting, a few years of clean history. For individual scoring, more than most organisations have, which is itself a reason for caution.
Should managers see flight risk scores?
Rarely, and only with clear guidance on what to do. A score without guidance changes behaviour unpredictably.
Is this regulated?
Automated decision making about individuals carries specific duties in several jurisdictions. Take advice before scoring people rather than after.