Hardware sales forecasting using clustering and machine learning approach

Abstract

This research is a case study of an information technology (IT) solution company. There is a problem that is quite crucial in the hardware sales strategy which makes it difficult for the company to predict the number of various items that will be sold and also causes the excess or shortage in hardware stocking. This research focuses on clustering to group various of items and forecast the number of items in each cluster using a machine learning approach. The methods used in clustering are k-means clustering, agglomerative hierarchical clustering (AHC), and gaussian mixture models (GMM), and the methods used in forecasting are autoregressive integrated moving average (ARIMA) and recurrent neural network-long short-term memory (RNN-LSTM). For clustering, k-means uses two attributes, namely "Quantity and Stock" as the best feature in this case study. Using these features the k-means obtain silhouette results of 0.91 and davies bouldin index (DBI) values of 0.34 consisting of 3 clusters. While for forecasting, RNN-LSTM is the best method, where it produces more cost savings than the ARIMA method. The percentage of the difference in saving costs between ARIMA and RNN-LSTM to the actual cost is 83%.