Modelo de elaboración de pronóstico de ventas mediante el uso de redes neuronales artificiales y SVR

Abstract

The importance of forecasting and predicting the level of sales is vital to the survival of a company. There is continuous research on the development for better systems due to the lack of clear sales patterns. Through the rapid advancement of technology, it has been possible to develop applications based on artificial neural networks (ANN). These applications focus on the development of sales forecasts of products of frequent consumption, improving the accuracy of traditional forecasting systems. In the present study, the performance of traditional models is compared with more developed systems such as neural networks and support vector machines or support vector regression (SVM-SVR). It demonstrates the importance of considering external factors as macroeconomic and internal indicators such as microeconomics, (for example, prices of related products), which affect the level of sales in an organization. These considerations had not been taken into account in the past. The application of this study was in a super-market. In the first instance, a preprocessor was carried out to clean, adapt and standardize the databases. Then, because there was no labeled information regarding which pairs of products were substitutes or complements, it was necessary to apply a cross-elasticity analysis. In addition, a harmonic average (f1-score) is considered in several points to establish priorities in some products and obtain results. Studies in which comparative analyzes are made between forecasts, which demonstrate the effectiveness of using neural networks is presented in: [7], [16], and [25]. The model proposed in this study presents a potential application in the forecast of sales of high turnover products in supermarkets, since their results are more accurate than those obtained with traditional procedures.