Valuation models, as currently implemented, tend to treat all model inputs and predictions with the same level of credibility; firms about which very little is known, for example, compete on the same basis with well researched companies. Analysts are, of course, aware of this problem and tend to correct for it by making adjustments to model inputs. In other cases, individual judgments may override model predictions. Such ad hoc judgments are subject to many of the same error bias problems as model predictions and tend to defeat the intended role of a valuation discipline.
Judgments about the reliability of model inputs must be made and incorporated in model predictions. What is needed is an estimate of how errors in judgment or model applicability are expected to contribute to the excess return implied by the valuation model. This can be obtained by determining the correlation of excess return predictions with the true information driving the return estimates. This measure of predictive reliability must be used to adjust raw model predictions in order to minimize error biases.
Because the reliability of information is generally not constant across the sample of securities being evaluated, model results must be scaled on a firm by firm basis. Differences in the level of reliability of a valuation model across a sample of securities are likely to have a strong impact on the model’s expected returns and rankings.
