A forecasting metric for predictive modeling

In science and engineering, simulation models calibrated against a limited number of experiments are commonly used to forecast at settings where experiments are unavailable, raising concerns about the unknown forecasting errors. Forecasting errors can be quantified and controlled by deploying statistical inference procedures, combined with an experimental campaign to improve the fidelity of a simulation model that is developed based on sound physics or engineering principles. This manuscript illustrates that the number of experiments required to reduce the forecasting errors to desired levels can be determined by focusing on the proposed forecasting metric.