Comparison Between Random Search and Genetic Algorithm in XGBoost Hyperparameter Tuning for Retail Sales Forecasting

Abstract

Sales is part of the important factor that influences a company in determining two things, namely profits and losses on the company. The right strategy to determine the amount of sales can be done through forecasting. Therefore, sales forecasting requires the right technique to produce accurate results. Machine learning has been proven to help overcome sales forecasting, one of which is XGBoost. However, XGBoost has many hyperparameters that affect its performance, it requires a hyperparameter setting method to produce an optimal hyperparameter. Random searches and genetic algorithms are optimized methods to find the optimal hyperparameter on XGBoost. The two methods of optimization were compared in this study with the measurement of RMSE performance in doing retail sales forecasting on the sales data of the retail company Rossmann Store which comes from the Kaggle site. The research obtained random search results superior to the genetic algorithm with RMSE values on the training process and the testing process are 0.123 and 0.122. Meanwhile, the RMSE values generated by genetic algorithms in the training and testing process are 0.333 and 0.332.