EXPERIMENTAL MODELING OF FEED FORWARD NEURAL NETWORKS FOR OUT OF STOCK PREDICTION: A COMPARATIVE HIDDEN LAYER APPROACH

Abstract

It is always not a pleasant experience when a customer comes to a commercial enterprise to purchase a product and that product is not available. To solve the problem, this paper presents an out-of-stock (OOS) prediction model. Achieving this aim involves methodologies such as data collection from OOS records, then data processing using standardized techniques such as imputation for replacement of missing values, Chi-square for feature selection, and principal component analysis (PCA) or feature transformation, respectively. The processed data were applied to train a neural network architectures varied with five different hidden layers, and the results obtained were evaluated and validated through comparative analysis. Findings showed the modeling of neural networks with increased depth of neurons has a great impact on the performance of OOS prediction models; however, one has to be careful in the modeling to address issues of overfitting.