Predictive Analytics For Sales Using Historical Transaction Data And Seasonal Trends

Abstract

The unpredictability of promotional and economic factors makes it imperative to have strong predictive models that would enable sound retail sales forecasting. The accurate sales forecasting in retailing is essential in managing inventory, minimizing operation cost, and maximizing customer satisfaction.  To enhance the accuracy of Walmart sales forecasting, the present research utilize a wide range of machine learning (ML) approaches supported by a large amount of data acquired in Kaggle. The methodology includes extensive data pre-processing, such as outlier removal, time-based feature engineering with the assistance of Exploratory Data Analysis (EDA). This is the reason an XGBoost regression model is developed because it has a high possibility to capture nonlinear relationships and can be used with large-scale data. High predictive accuracy is reflected in the results of the experiment with the best model performance as the R 2 of 0.99, Mean Absolute Error (MAE) of 1226.47 and the root mean squared error (RMSE) of 1700.98. A comparative study also proves that XGBoost is better than the classical models such as the Gradient Boosting, Decision Trees and the Random Forests. The results prove the efficiency of the suggested method and indicate its usefulness in practice when applied to retail forecasting.