Dynamic Quality Ladder Model Predictions in Nonrandom Holdout Samples

In light of recent calls for further validation of structural models, this paper evaluates the popular dynamic quality ladder (DQL) model using a nonrandom holdout approach. The model is used to predict data following a regime shift—that is, a change in the environment that produced the estimation data. The prediction performance is evaluated relative to a benchmark vector autoregression (VAR) model across three automotive categories and multiple prediction horizons. Whereas the VAR model performs better in all scenarios in the compact car category, the DQL model tends to perform better on multiple-year horizons in both the midsize car and full-size pickup categories. A supplementary data analysis suggests that DQL model performance in the nonrandom holdout prediction task is better in categories that are more affected by the regime shift, helping to validate the usefulness of the dynamic structural model for making predictions after policy changes.

This paper was accepted by Matthew Shum, marketing.