Normalized nonconformity measures for automated valuation models

Automated valuation models are used by the real estate industry and financial institutions to provide estimated property values and automate the time-consuming process of property valuation. We apply conformal predictors to the problem of automated valuations to provide prediction intervals of estimated price that are guaranteed to be reliable at a confidence level set by the user, just assuming the data is independently and identically distributed. In this study, we use the rich Ames data set of house characteristics and sale prices over 2006 to 2010 and explore alternative nonconformity measures, used as part of the conformal predictors, based on normalization techniques and model averaging. We find that we can produce efficient prediction intervals using these methods that are competitive with existing commercial automated valuation models and have the additional benefit of guaranteed reliability.