Forecasting Models Using Deep Learning Algorithms by Means of Recurrent Neural Networks LSTM to Estimate Demand of Building Making Materials

Resumen

Abstract– Forecasting is an area of knowledge of great interest to companies because not only it allows them to carry out their strategic planning processes but also to obtain valuable information from inventories of finished products, raw materials, materials and supplies to make purchasing and production decisions in the best way possible. Moreover, traditional mathematical forecasting models not always provide robust forecasting results, especially when the variability of historical data is significantly high. The objective of this work is to provide a methodology to forecast the demand of a family of products for building making by using deep learning algorithms through neural networks and compare the results with those obtained by a forecast model with automatic detection of seasonality and a forecasting model smoothed by third degree polynomial regression. The results obtained indicated that the forecast model using neural networks provided an RMSE forecast error 79.14% lower than that detected by the Holt-Winters model found with automatically detected seasonality. Similarly, the neural network forecasting model found a forecast error 79.76% lower than that detected by a third order smoothed polynomial forecasting model. The recommendations based on the results indicate that it is convenient to use neural network algorithms when the demand historical data presents significant variability and traditional forecasting methods explain very little of the data total variability.