Definition
Mean absolute percentage error between AI-forecasted car demand by type and region and actual car demand that materialized, measured at the positioning decision horizon.
The Measurement Gap
AI demand forecasting models tell car management where to position empty equipment, which car types, which regions, how many. But no legacy metric measures whether the forecast was right. Car fleet utilization measures how efficiently the fleet is used but cannot distinguish between cars positioned correctly (meeting demand) and cars positioned incorrectly (sitting idle where they were not needed, absent where they were). Equipment velocity measures how fast cars move through the network but cannot see whether they moved to the right place. OTP measures train schedule adherence but has no visibility into whether the cars on those trains were needed at the destination. The failure mode is a demand model that consistently mispositions equipment (sending gondolas to regions that need hoppers, flooding one area with cars while another faces shortages) at a cost the railroad bears in repositioning expense, lost revenue, and customer dissatisfaction, all invisible to existing metrics because no metric compares forecast to outcome at the car-type-and-region level.
Decision Supported:
Is the demand model accurate enough to justify the car repositioning decisions being made on its predictions? Which car types and regions have the worst forecast accuracy?
Formula
Plain language: For each car type in each region, compare the number of cars the model said would be needed at the forecast horizon to the number actually ordered by customers. Average the error across all car-type/region combinations.
Notation:
ECPA = 1 − (1/N) × Σ |Forecast_demand_cr − Actual_demand_cr| / Actual_demand_cr
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Forecast_demand_cr | AI-predicted car demand for car type c in region r at the forecast horizon | Cars |
| Actual_demand_cr | Actual car orders received for car type c in region r at the materialization date | Cars |
| N | Number of car-type × region combinations evaluated | Count |
ECPA is a percentage where 100% equals perfect accuracy. Higher is better. Companion metric: Repositioning Waste (RPW): the dollar cost of moving cars to places they were not needed.
Proposed formula, pending validation pilots.