What is a predictive attrition model?

It looks at the people who left in the past, finds the patterns that came before their exits, and scores current employees against those patterns. The output is a ranked list, or a risk level per person, for a coming period.

It is the systematic version of a flight risk assessment: the same question, asked of everyone at once and refreshed as the data changes.

Internal models use data from your HRIS and survey tools: tenure and time since the last promotion, pay against the role's range, manager changes, engagement scores and absence. Signal-based models watch public changes instead: profile headline and title changes, new skills or certifications, an Open to Work badge, and spikes in profile-edit activity. Internal data changes slowly. Public signals can move within a week.

A model ranks attention, it does not make decisions. Treat the output as a queue for manager conversations, and never as an input to pay, promotion or performance decisions.

Take an invented team of 40. The model flags 4 people as high risk this month. The manager reviews the reasons, holds 4 conversations about growth and workload, and learns that 2 of the 4 were already talking to other employers. Without the model, those conversations would have gone to whoever asked first.

The first check is how many of the people it flagged went on to leave. The second is how many of the people who left had been flagged first.

Take an invented year in which a model flagged 10 people and 4 of them left within six months: 4 ÷ 10 × 100 = 40% of its flags were right. In the same period 8 people left in total, so the model had flagged 4 ÷ 8 × 100 = 50% of the leavers in advance. Both numbers matter. A model that flags everyone catches every leaver and helps nobody.

Remember that a good retention conversation changes the outcome. Someone who was flagged and then stayed may be the model working, not failing.

• Using the score in pay, promotion or performance decisions.
• Trusting a score without looking at the reasons behind it.
• Training on too few past exits, so the patterns are mostly noise.
• Building a model and never acting on the list. A score nobody follows up changes nothing.

TeamPredict gives each employee you track a resignation-risk level from public LinkedIn signals, checked every day, with the specific profile changes behind it. It needs no HRIS integration and never contacts the people you track. See how it works.