This study examines the classification and forecasting of e- commerce orders in Pakistan using the Pakistan's Largest E- Commerce Dataset, with the objective of improving predictive accuracy and supporting data-driven decision-making in the digital commerce sector. For classification tasks, Random Forest and XGBoost were applied to predict binary order status (successful vs. unsuccessful). Random Forest outperformed XGBoost across all evaluation metrics, achieving 67.9% accuracy, 67.03% balanced accuracy, 54.2% sensitivity, 79.85% specificity, 70.12% precision, and an F1-score of 0.727, compared to XGBoost's 66.19% accuracy and F1-score of 0.709. Feature importance analysis revealed that grand total, price, and discount amount were the most influential predictors, underscoring the strong role of pricing strategies and promotions in determining order completion. For forecasting, ARIMA provided limited performance with a MAPE of 41.1%, while SARIMA enhanced seasonal accuracy (MAPE ˜ 21%). Prophet achieved the best forecasting accuracy (MAPE = 17.4%), effectively capturing both long-term trends and seasonal variations in sales. Taken together, the findings confirm Random Forest as the most robust model for classification, Prophet as the most effective model for forecasting, and SARIMA as a stronger alternative to ARIMA. This research contributes to Pakistan's e-commerce literature by demonstrating how advanced machine learning and time series methods can uncover key drivers of consumer behavior, highlight the importance of pricing, discounts, and seasonality, and provide actionable insights for improving customer retention, demand forecasting, and long-term digital growth.