Construction of predictive model of employees turnover

Abstract

Currently, there is a growing need for companies to manage their workforce, aiming to maintain qualified professionals and reduce costs associated with dismissal processes. Other than that, there are advances in the field of investigation of Machine Learning, which enables the description of future scenarios based on data-oriented predictive models. This combination of factors has enabled companies to invest in ways to predict when their employees are most likely to leave organizations, anticipating the loss of talent and reducing operating costs. Thus, this study aimed to build a predictive model of employee termination for a financial institution in Brazil, in addition to understanding the main factors linked to turnover. The study was conducted by testing the performance of the K-Nearest Neighbour, Multiple Regression, Naive Bayes and Random Forest algorithms in a database containing information from workers, collected over a year. It was evident that the best predictive model was built using the Random Forest technique, which presented an accuracy of 78.3% and a precision of 81.5%. It was also observed that personal characteristics, such as age and number of children, and professional characteristics, such as remuneration and annual performance evaluation, were the most relevant variables for classifying a professional as prone or not to leave the company.