Performance Analysis of Machine Learning Models for Sales Forecast

Abstract

Many supermarkets today do not have a strong forecast of their yearly sales. This is mostly due to the lack of the skills, resources and knowledge to make sales estimation. At best, most supermarket and chain store use adhoc tools and processes to analyze and predict sales for the coming year. The use of traditional statistical method to forecast supermarket sales has met a lot of challenges unaddressed and mostly results in the creation of predictive models that perform poorly.  The era of big data coupled with access to massive compute power has made machine learning model the best for sales forecast. In this paper, we investigated the forecasting of sales with three machine learning algorithms and compare their predictive ability. Three different methods used are K-Nearest Neighbor, Gradient Boosting and Random forest. The data used to train the machine learning models are data provided by Data Science Nigeria on the Zindi platform, the data were collected from a supermarket chain called “Chukwudi Supermarkets”. The results show that the Random Forest algorithm performs slightly better than the other two models, we saw that Gradient Boosting models were prone to over-fitting easily and that K-Nearest Neighbor even though fast, performs poorest among the three.