Today’s job market for highly skilled employees is white hot. The workforce is quick to move from one opportunity to another, and employers are faced with high costs for recruiting and reduced efficiencies resulting from having to constantly train new employees.

The exact cost of turnover is hard to estimate, since it varies significantly among different studies and situations. An oft-cited estimate based on a calculation using a “Cost of Turnover” worksheet provided by the Society for Human Resource Management (SHRM) is that costs are roughly 150% of the employee’s salary. Other calculations suggest that it is more than 200% for such positions as managerial and sales jobs. At the other end of the spectrum, alternative calculations by Boushey and Glynn estimate the average cost of replacing an employee to be roughly 21% of an employee’s annual salary.

While employee turnover, and its related costs, is the primary responsibility of human resource functions, CFOs in their role of overall expense management should be interested in this area. More importantly, CFOs could potentially save their companies a lot of money by reducing employee costs. How can they do that?

By using predictive analytical tools to interpret Big Data, thereby enabling management to take mitigating actions which can reduce this churn. Most companies already store vast amounts of data on their employees, such as employee personal profiles, job profiles, career progression information, demographics, compensation, stress / burnout, working conditions, and job satisfaction. By analyzing these data points, managers can come to understand the root causes of employee churn. Once this is understood, companies can take steps to mitigate the factors leading to increased employee churn, reducing turnover costs significantly.

The underlying reasons for churn include dissatisfaction with: the company, location, compensation, job, performance standards, and relationships with the supervisor and teammates. The challenge for most companies is to collect the data on all of these possible factors, and then evaluate which are the most highly linked to employee churn. A large number of mathematical and machine-learning algorithms can be used to model this problem and predict employee churn well in advance of the actual churn event.

In one recent case, Protiviti consultants used logistic regression to predict the probability of churn. The final model produced an equation that quantified how strongly each predictive variable influences the probability of an employee leaving, as well as a measurement of the statistical significance of the variables.

Information such as the employee’s college major, highest degree type, number of years at previous employers, history of change in salary, industry experience, and promotion history were found to be highly predictive of employee turnover in this instance. The model is not only useful for diagnosing employee turnover in the present, but also can help predict which employee is likely to leave in the future. By using the model, managers can determine how likely an employee is to leave, and the major driving factors behind the employee’s decision. Once these driving factors have been identified, the organization can take remedial action to reduce employee churn, which can lead to substantial savings, happier and more engaged employees, and increased productivity.

Such new tools are emerging on the market as SAP’s Predictive Analytics leveraging HANA and Lumira, as well as IBM solutions including various SPSS tools. These tools have begun to focus more on business users, putting the ability to gather, model, and analyze large data sets directly into the hands of end users. Software vendors are also working to better integrate their products with existing modeling tool sets.

Adding unstructured data sources would give modelers even greater insights into employee behavior and proclivities. Digital social networks like Twitter, Facebook, Glassdoor.com, Vault.com, internal messaging, WeChat, are WhatsApp are generally used by employees to vent, complain, and comment about their work environment.

The digital data from these sources can be harnessed to enhance predictive models, especially for Gen-X, Gen-Y, and Millennial employees whose behavior in the digital social media space may be predictive of his or her behavior in the job setting.

Developing a churn model by using both types of data – an organization’s internal employee data as well as social network data – can enhance the model and improve its predictive accuracy, thus leading to even more savings through better information for decision-making.

A word of caution about using unstructured data, however. While such data sources have predictive power, and many companies are already using them (Amazon sending targeted ads based on previous buying patterns, for instance, or Netflix suggesting movies consumers may like based on their previous viewing history), employers need to be careful about violating of privacy laws. Unintended consequences can arise out of actions against minorities and protected classes, which could lead to legal problems.

Shaheen Dil is a managing director at Protiviti.