Previsão de preços e demandas de produtos do varejo utilizando técnicas de aprendizado de máquina

• Autor(a) principal: Souza, Thiago Mantuani de
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/46869
• topic: Aprendizado de máquinas; Redes neurais artificiais; Máquinas de vetores de suporte; Algorítmo K-vizinhos mais próximos; Floresta Aleatória; Artificial neural networks; K Nearest neighbors; Machine learning; Random Forest; Support vector machines; Engenharia de Software

The success of retail companies depends on some factors that help in decision making. One of these factors is related to the storage and availability of products, in order to meet customer demand. Prices are also one of these factors because, based on them, customers will make the decision to purchase the products. Thus, the objective of this work was to apply machine learning (ML) techniques to predict the demands and prices of some retail products. For the ML system training, a series of retail sales of some products was chosen, covering the period from April/2015 to December/2019 in the city of Cambuí/MG. The ML techniques applied and compared were: Linear Regression, Multilayer Perceptron Artificial Neural Network, Long Short Term Memory Recurrent Neural Network, Support Vector Machines, K Nearest Neighbors and Random Forest (RF). The results of demand and price forecasts were obtained through daily sales and evaluated through the metrics of the root mean square error (RMSE), root mean square logarithmic error (RMSLE), mean absolute error (MAE) and coefficient of determination (R²). After the execution of the ML models referring to thirteen different periods, the RMSE, RMSLE, MAE and R² of each of these periods were obtained. Subsequently, Friedman's non-parametric test was applied to verify whether there was a statistical difference between the means and the Nemenyi test to identify which models were different. The RF model provided the best predictions for retail product prices and demands. In this case, the values calculated for the RMSE, RMSLE, MAE and R² metrics, through the RF for price forecasting, were close to 0.07, 0.03, 0.11 cents and 0.99 respectively. In the demand forecast when the RF algorithm was applied, the calculated value for the RMSE was approximately 55.6, while the calculated RMSLE value was 0.63 and the MAE was close to 4 product units. Finally, the value found for R² was 0.57. Thus, RF proved to be an efficient method for forecasting prices and demand for retail products covered in this work.