Um modelo logístico bayesiano para estimar a probabilidade de desligamento de vendedores do setor de telemarketing de uma empresa

Abstract

In the telemarketing sector of a wholesaler company, it is constantly dealing with high turnover rates, resulting in high costs whith remediation, training, selective processes, among others. Considering that employees usually return, after a year worked, the investment made for them, the problem is treated in terms of probabilty, in wich the idea is to predict probabilistically when a saller is prone to departure. For that, a logistic model was used, treating as an answer event when the telemarketers are dismissed with less than a year worked. 19 covariates we considered in the analysis. The model parameters were estimated using Bayesian inference. In this methodology the previous knowledge is updated using the information derived from the data, allocated in the likelihood. It is assumed that prior knowledge about the parameters is described by probability distributions. The previous information of the data is translated into an a posteriori distribution through the Bayes' theorem. As the posterior distribution, in general, has a complex form, samples of the complete conditional distributions are obtained for each parameter iteratively and recursively using algorithms that makes use of the Monte Carlo method by Markov chain, when the algorithm reaches convergence, the Conditional samples represent satisfactorily and empirically the posterior distributions. The conversion study can be done using the Geweke criterion (Geweke, 1992) and Raftery and Lewis criterion (RAFTERY \& LEWIS, 1992).To select the important variables in the occurrence of the event of interest, it was based on the genuinely Bayesian hypothesis test (FBST) and its e-value.The procedure for selecting variables was similar to the backward in classical inference. From the fitted model it is possible to predict the probability of a seller being disabled, being necessary for this, only to know their profile based on age, average billing and time in which they remain in call, since these were the selected variables to compose the final model. The assertiveness of the model was 73% and therefore, based on this criterion, the adjustment was considered adequate.