Forecasting grocery item sales using gradient boosting models: A study of GridSearchCV, RandomizedSearchCV, and optuna optimization approaches

This study introduces predictive modeling based on Extreme Gradient Boosting (XGBoost), which utilizes Optuna for hyperparameter optimization and evaluates performance against GridSearchCV and RandomizedSearchCV using a 500-day dataset. To ensure statistical reliability of the findings, a bootstrapping with 1000 iterations is used to calculate 95% confidence intervals for all performance measures. Although GridSearchCV and RandomizedSearchCV achieve consistent performance, their average R-squared test performances are 0.81677 and 0.85538, respectively. In contrast, Optuna outperforms both methods by identifying better regions of optimal parameters, with an average R-squared of 0.94146. Furthermore, the computational efficiency analysis shows that Optuna’s average execution time of 28.31 s is a practical trade-off for running on consumer-grade hardware. Model interpretability is confirmed through Shapley Additive Explanations (SHAP) analysis, which identified the 3-day rolling average as the most important driving factor of demand.