With all the negatives associated with high employee turnover, wouldn’t it be great if companies had a crystal ball that would show them which employees are likely to leave?

Actually, there is a way to determine which employees are more likely to leave, but it’s science-based, not magic: Predictive modeling uses historical data to glean insights into which employees are more likely to seek greener pastures.

By compiling key employee data — often already on file — and running it through such a model, organizations can then take steps to try to keep in the fold those employees identified as high flight-risks. By doing so, they can minimize negatives like high recruiting costs and low morale, says Jacque Carlson, a data scientist for BlueGranite, a data and analytics consulting firm.

“Predictive models offer another way you can reduce costs associated with things like advertising for job openings, the time it takes to do interviews, onboarding costs and so on,” she explains. “Then there’s the low morale that high turnover creates. When the rats are fleeing the ship, those that stay behind start wondering if they should leave, too.”

Using predictive models comes with some caveats, however. First of all, it’s important for organizations to understand that different categories of employees require different data models; there’s no one-size-fits-all model for an entire company.

“All career paths are qualitatively different,” Carlson says. “You need to create different models depending on the group on which you’re focusing.” 

In addition, the accuracy of predictive models depends on the integrity of the data upon which they’re built. Or as Carlson puts it, output equals input.

“The model is only as accurate and unbiased as the data fed into it,” she says.

The breadth of data matters, too. A good predictive model needs at least two to three years of data to provide accurate results. Moreover, a predictive model is not the ultimate arbiter in reducing turnover. Instead, it should serve as just one of a collection of tools aimed at divining what drives employee turnover.

After data has been collected and analyzed, it’s important to recognize that the numbers alone don’t tell the whole story. To make the most of the data collected, it’s incumbent on organizations to probe behind the numbers. That means talking to employees tagged as high-flight risks; during that discussion, you must be able to explain how that was determined.

“They’ll ask you why you think they’re a flight risk,” Carlson says. “And if you can’t answer that question, the model has no value.”

That raises another important point: It’s critical for organizations to let employees know it’s collecting data and to give them the opportunity to opt out if that makes them uncomfortable.

“Transparency is the important point. It’s the first thing I talk about with clients,” Carlson says. “In terms of ethics, collecting data for models about human beings is qualitatively different than collecting data about, say, a machine that makes steel. The machine will never care about that, but employees sure might.” 

So what kind of data should organizations collect that can indicate employees are flight risks? The list is long, but here are a few examples, some obvious and some not so much:

• How far they commute, especially if bad weather strongly impacts the quality of commute for long periods of time.

• The number of supervisors an employee has had (a revolving door of managers may create disjointed career paths for their direct reports, which spurs their departure).

• How many times they’ve been promoted or received a raise, as well as how long it’s been since their last promotion.

• How much personal time off employees use compared to what’s allotted; employees that receive, say, two weeks of PTO annually and rarely use it are good candidates for burnout — and flight.

• Other demographic data, such as gender, age, ethnicity, and education level, along with factors such as whether or not they’re parents, homeowners, and divorced. This information can be gathered in a variety of ways, including having employees periodically fill out forms to update information.

How accurate are these predictive models? While predictive modeling is very complex to explain, in the end, a data scientist compares the results produced by the model to acceptable standards of something called evaluation metrics.

“Most models are rejected if the accuracy is less than 90 percent,” Carlson says. “However, if the data is good quality and there is more than a sufficient amount of data, the standards for the model should be quite high. The evaluation metrics should all be in the high 90s.”

Given the complexities of predictive modeling and given that data science is a relatively young field, most companies and organizations aren’t capable of building flight-risk models. But management or human resources leaders can self-educate themselves and compile data, then hire a data scientist to build a predictive model. This might or might not be feasible and practical depending on the size of your operation, but it’s good to know it’s an option.

What companies do after that, in terms of how they go about convincing flight-risk candidates to stay, is not within the data scientist’s purview. But one thing is certain: With a shrinking labor pool creating intense competition for new employees, there’s definitely value in identifying those departure-minded employees and trying to keep them onboard.

“Ignoring the value of predictive models can be companies’ downfall,” Carlson says. “The models can provide great insights into things happening in organizations that drive the risks of leaving.”