Strengths and Limitations of the Predictive Models

Understand the accuracy of Visier's predictive models and the data limitations that affect their results.

Overview

Visier's risk of resignation model has been shown to identify likely resignations up to 17 times more accurately than guesswork, with an average of 5 times more accurate across clients. Prediction accuracy varies by customer, largely driven by how much data is available and by industry- or location-specific patterns in employee turnover.

Several factors also limit what the models can use. Properties with more than half their values missing are automatically excluded. Sensitive data, like age or gender, and backdated data that reflects future outcomes should be excluded to avoid skewing results or reinforcing bias. Training data should also be scoped to the population you're actually interested in, such as limiting a resignation model to permanent employees only.

Strengths

Visier's data scientists are working continuously to improve the predictive strength of the machine learning algorithm for all customers. We measured the predictive success of the risk of resignation model by looking at the predictions we made for all our clients. For each client, we looked at the employees who we predicted as having the highest likelihood of resigning and determined how many of them actually resigned in the following year. The results show that Visier's risk of resignation predictions can be 17 times more accurate than guesswork. The average predictive success for all clients was measured at 5 times more accurate than guesswork.

The reasons why some of our customers do much better than the rest is due to:

• Data volume: The more information that is available the better the predictions.
• Customer specific differences: Some industries and locations have a more predictable employee turnover behavior.

Limitations

Properties with more than 50% missing values are excluded from the predictive models.

As a best practice, we recommend excluding sensitive data and backdated data. Sensitive data, such as an employee's age or gender, may inadvertently reinforce subconscious biases within your organization's recruitment or retention. Backdated data, or data that has information about future events, may negatively impact the accuracy of your predictions because you are including information in the predictive algorithm that was not actually available.

Additionally, your training data should only include populations that you are interested in. For example, your Predicted Risk of Resignation model should only include permanent employees. Employee populations can be filtered using the subject filter in the Customize tab. For more information, see Configure the Predictive Models.