Optimization of XGBoost Hyperparameters Using Three-Dimensional Learning AVOA for Retail Demand Prediction

Abstrak

Accurate demand forecasting is critical for retail supply chains, particularly in the Fast-Moving Consumer Goods (FMCG) sector, where even small discrepancies between predicted and actual demand can lead to excess inventory or stock shortages. This study proposes a hybrid TDLAVOA–XGBoost model that adaptively optimizes key hyperparameters to improve forecasting accuracy and stability. The analysis is conducted using 990 FMCG inventory records from a publicly available dataset to examine the impact of metaheuristic-based optimization on model performance. The TDLAVOA algorithm identifies an effective hyperparameter configuration (max_depth = 3, learning_rate = 0.01, n_estimators = 100, gamma = 1.97, subsample = 0.57, and colsample_bytree = 0.66), enabling the proposed model to achieve an RMSE of 22.53 ± 0.50 and an MAE of 19.32 ± 0.33. Compared with the default XGBoost baseline, this represents a substantial reduction in prediction error and variability. Comparative results show that TDLAVOA–XGBoost achieves performance comparable to SARIMAX and demonstrates superior accuracy relative to deep learning models, including LSTM and MLP, for limited-sample tabular FMCG demand data. Statistical validation using one-way ANOVA and Tukey’s HSD confirms that the performance differences among models are statistically significant (p < 0.0001). Overall, the findings indicate that TDLAVOA–XGBoost provides a practical and reliable approach for supporting data-driven inventory planning in retail environments.