Eventually all large transformers will be dynamically loaded using models updated regularly from field measured data. Models obtained from measured data give more accurate results than models based on transformer heat-run tests and can be easily generated using data already routinely monitored. The only significant challenge to using these models is to assess their reliability and to improve it as much as possible. In this work, we use data-quality control and data-set screening to show that model reliability can be increased by about 50% while decreasing model prediction error. These results are obtained for a linear model. We expect similar results for the nonlinear models currently being explored.