Forecast accuracy versus operational trigger value in next-day truck utilization forecasting: Evidence from Philippine logistics

This study examines whether one-day-ahead predictive analytics can support fleet-planning decisions in an anonymized Philippine logistics operation and whether conventional forecast accuracy corresponds to operational trigger value. Using audited daily records from 2024-2025, the study constructs a temporally validated forecasting panel for realized served-truck utilization, defined as served deliveries per available truck. Because this outcome is constrained by demand, capacity, readiness, and execution, supplementary demand-pressure targets are constructed from scheduled deliveries, open orders, deferred or rescheduled work, and backlog-adjusted pressure. The analysis benchmarks short-memory baselines, linear models, Random Forest, XGBoost, LightGBM, and an LSTM with a 28-observation lookback window under expanding-window validation and an untouched 2025 holdout. After multiplicity adjustment, the moving-average baseline had significantly lower paired absolute error than the LSTM benchmark, but did not differ significantly from Random Forest, XGBoost, or LightGBM. The findings show that daily logistics forecasting should select models by planning use case: conservative next-day estimation, demand-pressure surveillance, or proactive intervention for stress and backlog risk.