Transformación digital en clientes de distribución : modelo predictivo de ventas con redes neuronales para automatizar la validación y monitoreo de ventas en consumo masivo

Abstract

This project develops a sales prediction model using machine learning techniques, specifically Recurrent Neural Networks (RNN) and SARIMA models, to automate the validation and monitoring of sales data from Procter & Gamble (P&G) distributors in the mass consumption channel in Colombia. Through an exhaustive exploratory analysis of historical sales data, key patterns, trends, and correlations in the time series are identified. Different neural network architectures, including Long Short-Term Memory (LSTM) and feedforward, as well as SARIMA models, are designed, trained, and compared to determine the most appropriate approach in terms of accuracy and generalization capability. The results demonstrate that LSTM networks outperform SARIMA and feedforward models, achieving a MAPE of 3.402%, an RMSE of 2.347 million, and a correlation of 0.987 for one analyzed client. However, for another client, the results are less satisfactory, suggesting the need to explore additional techniques according to the specific characteristics of each series. The implementation of these predictive models can generate benefits for P&G, such as early detection of anomalies in distributor reports, optimization of commercial strategies, and decision-making based on reliable and timely information throughout the mass consumption supply chain.