Resumo:
The recurring purchase business model, such as subscription clubs, has
grown significantly in recent years. Since acquiring new customers is more costly
than retaining existing ones, customer loss (churn) negatively impacts business
competitiveness, making the ability to predict it a strategic advantage. The variety
of available methods and the lack of a standardized model evaluation process
pose both academic and practical challenges for the sector. In this context, this
study structured a procedure for developing churn prediction models, consisting
of three stages: data preparation, model application, and model evaluation. First,
the data is standardized and balanced, comparison metrics are defined, and lastly
a stability analysis is conducted. To validate the method, its application was
performed in a real case of a book subscription club. Seven predictive variables
were selected to train three models: logistic regression, a combination of the kmeans
algorithm with logistic regression, and a multilayer perceptron (MLP)
neural network. Logistic regression was identified as the best-performing method,
achieving 71,1% of accuracy and 44,6% of precision. These results are superior
to the actual model used by the company, that has 71,8% of accuracy and 12,1%
of precision. Overall, the structured procedure proved effective in model selection
and decision-making for churn prediction development in a standardized and
robust way.
Keywords:
