Finding
On 184 test days, the selected ridge model has weighted absolute percentage error of 13.35%, compared with 18.70% for a seven-day seasonal baseline. Its nominal 90% prediction band covers 87.5% of observations, below the intended level.
Why it matters
Compare forecasting methods using both error and the consequences of underplanning or overplanning. The capacity-cost examples depend on assumed costs and do not measure operational savings.
Who could use this
An operations planner deciding how a rental forecast should influence a capacity rule.
Data and scope
731 daily observations from the UCI Bike Sharing dataset, covering historical 2011–2012 Capital Bikeshare rentals. Capacity-cost inputs are hypothetical.
How the analysis works
Five forecast specifications, chronological validation, rolling one-day-ahead predictions and a separate 184-day test period. Validation selects ridge regression with penalty 1. Residuals inform asymmetric-cost planning rules.
Limits of the evidence
The target is realized rentals, not uncensored demand. Trip-equivalent capacity is not a physical fleet-size recommendation. Any modeled cost difference is a counterfactual under invented costs, not observed savings.