SAP SuccessFactors Workforce Analytics & Planning (WFA&P) provides a set of analytical and planning capabilities to simplify accessing data from multiple and disparate sources for Human Capital Management (HCM). WFA&P consists of Embedded Intelligence (Report Designer, Ad-Hoc Reporting, etc.), Workforce Analytics (Headlines, Benchmarks, etc.), and Workforce Planning (Forecasting, Risk Analysis, etc.). It can answer questions such as “Is turnover an issue?”, “What type of employee tend to leave voluntarily?”, and “Who are the top 10 highest impact-of-loss employees would leave?”
WFA&P is working on extending its analytical and planning capabilities with predictive analytics to provide additional insights for various HCM topics. The predictive analytics capabilities will add the “Why” and “How” facets to the existing analytics capabilities, such as the answers to the questions “What are the key drivers for turnover?”, and “What will make those employees stay?”
Employee Flight Risk
One of the hot topics is that of voluntary terminations (or employee flight risk), when employees decide to leave the company on their own. Like customer churn prediction, which predicts the likelihood of a customer to cancel a service offered by a business, employee flight risk predicts the likelihood of an employee leaving a company voluntarily, by comparing and contrasting the employee details for those of who left on their own, with the employee details for those who stayed.
Flight risk prediction utilizes two sources of information – employee headcount and voluntary terminations. Employee headcount is the actual number of people employed at the end of a reporting period and voluntary terminations represents the number of employees who terminated their employment voluntarily with the organization.
Employees details are represented by a set of workforce dimensions, which provides a 360-degree view of an employee. The dimensions are organized in categories, such as demographics (age, gender, disability, ethnicity, etc.), compensation (salary, stock options, etc.), development (key position, performance rating, potential rating, etc.), employment (job category, employee class, employment level, grade, etc.), succession (critical job role, succession rating, successor readiness, etc.), and tenure (grade tenure, organization tenure, position tenure, time in grade, etc.)
Automated Predictive Modeling
Using SAP Automated Predictive Library (APL) for its predictive algorithms and data mining environment is one approach to address flight risk predictions. APL is based on the concepts of VC dimension and Structural Risk Minimization to keep the model simple and robust. It employs ridge regression, a non-parametric algorithm, to minimize the need for making assumption for data distributions. Its data mining environment facilitates automatic parameters tuning and model selection among other things, to free the users from performing these tasks manually.
The process of extracting historical headcount data to create a model and apply that model to the current employee headcount data has been automated and is shown in Figure 1.
Figure 1. Diagram illustrating the steps of an automated approach for creating and applying flight risk prediction with associated input data and model output.
Prediction Performance
Many factors contribute to the quality of a predictive model. Two of the basic factors are attribute size and record size. If the number of attributes is small, then the model is likely to be biased. Further, if the number of records is small, the variance may be too high to separate one outcome from another. Therefore, having an adequate number of both attributes and records are essential to creating a high quality predictive model.
During the development, we have examined the relationship between the number of attributes used and the prediction performance in terms of predictive power, which translates to prediction accuracy: a ratio of correct predictions over the total number of records. We found that the predictive power is consistent with the number of attributes used in creating the model.
Data used for this analysis come from WFA&P customers in various industries, organized by the workforce categories described earlier. As shown in Figure 2, the available attributes and records from these customers differ widely. In each cell, the ratio, which ranges from 0 to 1, represents the amount of attribute available for a given attribute category. A 1.00 means the customer has all the attributes in that category, and a 0.0 means the customer has no attribute in that category. For example, customer 1 has all the demographics and compensation attributes, but it does not have any Succession attribute.
Figure 2. Heatmap of ratios representing amount of attribute used by category among customers.
As expected, the quality of the models is affecting by the number of attributes used for creating predictive model. As shown in Figure 3, the predictive power represented by the color and shade, is largely consistent with the number of attributes used to creating the model.
Figure 3. Heatmap of comparison showing consistency between attribute ratio and predictive power.
Prediction Insights
Besides model quality, another important quality measure of a predictive analytics application is its ability to provide insights that are easy to understand. In APL, the insights are represented as Influencers. As shown in Figure 4, the influencers of a predictive model for a given sample customer include grade (“Grade Band”), organization tenure, performance rating, and job function (“Function View L3”). Among them, grade has the greatest effect on employees’ decision to leave, followed by organization tenure, and so on.
Figure 4. Bar chart of influencers showing relative prediction contribution.
These insights can be further refined by examining the categories of each influencer, where categories are the values a given influencer can have. For example, the categories for performance rating include “High Performer”, “Mid Performer”, “Low Performer”, and “Not Rated”. As shown in Figure 5, the categories “Low Performer” and “Not Rated” have a positive effect on flight risk; whereas, the categories “Mid Performer” or “High Performer” have a negative effect on flight risk. In other words, employees who are low performer are more likely to leave; whereas, employees who are mid-performer or high performer are less likely to leave.
Figure 5. Bar char of influencer categories showing relative positive and negative influences.
Finally, the hallmark of a predictive application is its ability to predict the future. Equipped with the insights in terms of influencers and associated categories, the application can generate a prediction for each employee in terms of risk of leaving and a set of drivers (or influencers) as the potential cause to leaving. This information can be used to mitigate the situation. For example, knowing that a high flight risk employee has a certain grade for a long period may trigger a compensation action to mitigate the situation.
Conclusion
Predictive analytics enables organizations with historical employee details to start predicting employee flight risk almost immediately with little effort. It can easily and quickly determine likely causes and problem areas. As a result, mitigation programs can be designed and implemented to address the attrition problem.
As an example, a customer was experiencing attrition of more than 50% in one year for new hires. After analyzing the problem with predictive analytics, one of the root causes was scheduling new employees with insufficient work hours since ramping up took longer than expected. As a result, the number of scheduled hours was increased to over 20 hours per week, especially in the first 90 days, for mitigating risk of attrition.
This example illustrates how using predictive analytics can discover unexpected and valuable insights. The predictive analytics approach will speed up the discovery and the resolution of HCM issues, compared to the traditional drill-down, and thus help address more key business problems more efficiently.
Acknowledgement
The author would like to thank Tony Ashton, Yvonne Baur, and Kouros Behzad for their insightful suggestions for this article.