Predicting Best with Imperfect Data

Abstract Anyone who plans uses prediction of some kind. However, the most elaborate and inclusive predictive model is not necessarily the best for applied work. This may be the case if the cumulation of data errors exceeds the predictive gain from superior specification, a stage which some of our most ambitious models may have reached. Rules of thumb for choosing and building models by this criterion suggest that, when complication leads to negative returns, a strategy of netting out simple, complementary models may be better. In general, poorer data call for simpler models. In passing, a technique is suggested for estimating the value of improvements in data. Meanwhile, elaborate models which are poor predictors may serve as useful contexts for partial models, and may achieve their full worth if maintained and improved over time. Further, even a poor predictor may contribute significantly to scientific knowledge and to the understanding of processes, and thus be helpful for making decisions. Some features of current practice may keep us from receiving the full benefit of the most advanced work in the modeling field.