Machine Learning Enhanced Demand Forecasting and Portfolio Analytics for Scalable Growth Management in Digital Logistics Operations

Abstract

This study developed and quantitatively evaluated a machine learning–enhanced demand forecasting and portfolio analytics framework for scalable growth management in digital logistics operations. A longitudinal analytical design was applied to 36 months of operational and commercial data comprising 128,560 eligible shipment transactions, 14,820 customer accounts, 46,735 bookings, 12 geographical markets, seven logistics service categories, 10 operating facilities, 214 vehicles, 3,684 completed routes, and 428 workforce records. Overall data completeness reached 96.4%. Mean monthly demand was 3,571.1 shipments with a coefficient of variation of 19.2%, while maximum monthly demand reached 5,214 shipments and demand increased by 18.7% across the observation period. Conventional statistical forecasting methods were compared with random forest, support vector regression, artificial neural networks, gradient boosting, LSTM, and XGBoost using chronological validation and independent testing. XGBoost achieved the strongest forecasting performance, with R² = 0.932, MAPE = 5.91%, MAE = 7.21, and RMSE = 10.18, representing a 54.5% reduction in MAE relative to SARIMA. Feature engineering increased XGBoost R² from 0.861 to 0.932. Portfolio segmentation demonstrated substantial commercial concentration, with the highest-value 20% of customers generating 63.7% of total revenue and 55.4% of shipment activity. Demand growth was strongly correlated with revenue growth (r = 0.812), while the multivariable commercial model explained 71.8% of adjusted revenue-growth variance. Capacity utilization averaged 82.4% but reached 97.3% during peak conditions, when fulfillment declined to 87.9%, overtime increased to 18.7%, and cost per shipment reached $89.70. A practical scalability threshold of approximately 90% utilization was identified. Statistical comparisons demonstrated large effects, including Cohen’s d = 1.31 for the XGBoost–SARIMA forecasting difference. The integrated framework explained 88.4% of subsequent portfolio-growth variation and classified scalable versus capacity-constrained growth with 89.7% accuracy and ROC-AUC = 0.941. The findings demonstrated that scalable digital logistics growth depended on integrating accurate machine learning forecasts with customer value, commercial performance, portfolio diversification, service reliability, and capacity availability.