Paper Title
ENHANCED SALE FORECASTING WITH TEMPORAL FEATURES AND HYPERPARAMETER TUNING

Abstract
For retail sectors to develop supply chain management and maximize inventory while maximizing customer stakeholder satisfactions, higher accuracy in sales forecasting is critical. The XGBoost Machine Learning Algorithm is used in this study to investigate gradual improvement towards accuracy of sales forecasting. Beginning with a basic model and taking a methodical approach-periodic model updating with various feature selections: for instance, rolling features, lag-based features, early stopping techniques, advanced hyper-parameter optimization methods like RandomizedSearchCV and Optuna. The dataset used is a cleaned supply internal historical sales data from Walmart, analysts used to include economic factors, store-specific sales, and holiday influences. Starting with an R2 of 0.48 in the baseline model, the study ended with an R2 increasing to 0.99 in the final optimized version, indicating significant model performance improvement. While systematic hyper-parameter optimization guaranteed the best model configurations, the incorporation of lag features and trendbased feature engineering displaced the model to better fit temporal dependencies. This thoroughly studies how critical feature engineering and strong methodologies of optimization are to achieve nearly-perfect accuracy in predicting. In retail and other related time-series forecasting conundrum, the results furnish a paradigm for the deployment of scalable, highperforming, predictive models. Keywords - Sales Forecasting, Predictive Modelling, Hyperparameter tuning, Feature Engineering