In verification circles, diagnostic measures are employed to convey to the user useful information about forecast errors. However, that information is not necessarily useful for a model builder. An example of what a model builder would welcome is guidance on how to set the the parameters in the underlying numerical model producing the forecasts. Guidance can be provided in two forms, motivated by the following two questions: 1) How does varying a set of model parameters, affect the resulting forecast parameters? 2) How should one set the model parameters in order to achieve some desired forecast parameters in an optimal sense? Both questions are difficult to address because of the nonlinear/interactive nature of the relationship between model parameters and forecast parameters. In this talk, multivariate statistical models (e.g., polynomial regression and neural networks) are utilized to capture both the "forward model" - mapping model parameters to forecast parameters - and the "inverse model" - mapping forecast parameters to model parameters. The numerical models used in the illustration are the 1963 Lorenz model, and COAMPS. The developed to capture that relationship. The utility of the approach is displayed by deriving results which would not be evident without the the statistical models.