Prediction on large scale data using extreme gradient boosting

Abstract

This paper presents a use case of data mining for sales forecasting in retail demand and sales prediction. In particular, the Extreme Gradient Boosting algorithm is used to design a prediction model to accurately estimate probable sales for retail outlets of a major European Pharmacy retailing company. The forecast of potential sales is based on a mixture of temporal and economical features including prior sales data, store promotions, retail competitors, school and state holidays, location and accessibility of the store as well as the time of year. The model building process was guided by common sense reasoning and by analytic knowledge discovered during data analysis and definitive conclusions were drawn. The performances of the XGBoost predictor were compared with those of more traditional regression algorithms like Linear Regression and Random Forest Regression. Findings not only reveal that the XGBoost algorithm outperforms the traditional modeling approaches with regard to prediction accuracy, but it also uncovers new knowledge that is hidden in data which help in building a more robust feature set and strengthen the sales prediction model.