Forecast Value Is a Property of the Decision: Calibrated Quantile Demand Forecasting for Urban Last-Mile Capacity and Routing

Forecasting studies routinely read lower prediction error as operational value. For service capacity planning that inference is unsafe: the cost-minimising commitment is a demand quantile whose level is the target service level, so a median forecast silently commits an operator to a 50% target. We test this over 263 New York City zones and 39.0 million taxi trips from two services with contrasting demand geographies. Spatio-temporal graph networks, a multi-relational graph-attention model and Graph WaveNet, are trained on seasonal residuals with a monotone quantile head under a pinball objective, then calibrated by split conformalised quantile regression on a later held-out block. The same forecasts feed two decisions that consume different functionals: newsvendor capacity commitment under a 9:1 underage-to-overage penalty, which needs a calibrated tail quantile, and top-K zone selection, which needs only a ranking. Evaluation uses chronological splits, multiple seeds and Diebold–Mariano tests. The two decisions give opposite answers. Committing the calibrated 0.9 quantile instead of the median lowers capacity cost by 42–51% and raises the achieved fill rate from 94.9% to 98.9% on one service and from 86.8% to 97.4% on the other, because the median is mis-specified for an asymmetric penalty rather than inaccurate; zone selection improves by under 0.15 percentage points over a weekly historical average. The lowest-error model is not the least costly, no architecture ranks stably across the two regimes, and no hand-designed graph relation replicates. The calibration layer transfers unchanged between backbones and reproduces 40–54% cost reductions on external taxi, parcel, food-delivery and courier corpora. The appropriate unit of evaluation is the forecast–decision pair.