Models are a simplification of reality, but all the parameters that relate a stress to system response need to be included to accurately predict the response to a future stress. Some parameters important to prediction accuracy may be lost in the simplification process. At the same time funding constraints make it difficult to decide how efforts should be divided between model construction and the collection of additional data. Modeling reports commonly conclude with a desire for more data and calibration; how well the model can simulate future stress, and how worthwhile the modeling effort is to decision makers, is often uncertain. To address these issues, a strategy for constructing models is proposed that uses regularized inversion and single-value decomposition. Using this approach, initial parameter complexity more closely reflects the underlying detail of the system and uncaptured detail is minimized. Moreover, model parsimony is employed automatically through single-value decomposition process resulting in a more stable and well-posed parameter estimation. In addition to non-linear regression benefits, the approach proposed can be used for a pre-calibration analysis that identifies: 1) parameters important for a prediction of interest; 2) an estimate of the commonly uncharacterized model structure error; and 3) the amount of parameterization needed to minimize the total prediction uncertainty. A case study using a steady-state model of the USGS Trout Lake Water, Energy and Biogeochemical Budgets (WEBB) site in northern Wisconsin is presented to illustrate the approach.