An Extreme Gradient Boosting Model Optimized with Genetic Algorithm for Sales Forecasting of Retail Stores

Forecasting has always been a curious topic to investigate for practitioners, academics and workers in private companies. Not only in the world but also In Turkey, COVID-19 pandemic makes difficult to forecast sales for any type of companies since patterns, sales and seasonality factors in sales have changed because of different reasons. At this point, the accuracy of sales forecasts is of great importance for retail companies. In particular, sales forecasts affect the decisions and actions taken on a daily and weekly basis. In this study, firstly, a model based on the Extreme Gradient Boosting (XGBoost) algorithm is proposed for daily sales forecasting of retail stores. Later, a hybrid GA-XGBoost model is developed to improve the performance of this model. In this model, the parameters of XGBoost are optimized by Genetic Algorithm. Finally, the performance of the developed model is compared with the SARIMA model using the root mean square error (RMSE), mean absolute percentage error (MAPE) and R-squared. The performance comparison is demonstrated by a case study with data from airport stores of a retail chain in Turkey. Numerical results show that the hybrid XGBoost-GA model outperforms the XGBoost and SARIMA models.