A Feature Engineering Approach for Tree-based Machine Learning Sales Forecast, Optimized by a Genetic Algorithm Based Sales Feature Framework

In the new retail era, the sales pace and consumers’ consuming habits are disrupted dramatically by e-commerce and mobile consuming market developments and become various, volatile, and complicated. In sales forecasting, tree-based machine learning algorithms are popular and impressive. Feature engineering, on the other hand, is the most significant impediment to blending business and technological knowledge. An innovative way to solving the feature engineering problem of employing tree-based algorithms in sales forecast is proposed in this work. And a genetic algorithm is presented to enlarge the feature base and to explore potential values from features. To ensure that features with important business characteristics are picked, a novel feature framework for sales forecast is offered. Also given is a genetic algorithm for expanding the feature base and exploring potential values from features. The performance of tree-based models from AutoGluon will next be evaluated using a public dataset, Kaggle Rossmann sales data. The results suggest that the proposed strategy can improve the accuracy and stability of decision tree algorithms in sales forecasting significantly.