Accurate demand forecasting is a critical challenge in modern retail supply chains, as it directly impacts inventory planning, cost optimization, and customer satisfaction. This study applies machine learning techniques to predict weekly sales using historical retail data. After preprocessing and exploratory data analysis, multiple models were evaluated, including Linear Regression, Random Forest, XGBoost, and LightGBM. Results showed that Linear Regression performed poorly with an RMSE of approximately 16,019, while tree-based models significantly improved accuracy. The tuned XGBoost model achieved the best performance with an RMSE of 4,552, representing a 72% improvement over the baseline. Feature importance analysis revealed that store ID, year, unemployment rate, and CPI were the most influential predictors of sales. Residual analysis confirmed stable predictions, and actual vs predicted plots demonstrated close alignment between forecasts and true values. While the model achieved strong results, limitations included restricted dataset features and computational requirements. Overall, this research demonstrates that machine learning, particularly XGBoost, provides an effective and scalable approach for demand forecasting in supply chains.