Combining Algorithmic and Stochastic Data Models

In this paper we combine a random effects model with different machine learning algorithms via an iterative process when predicting commercial real estate asset values. Using both random effects and machine learning allows us to combine the strengths of both approaches. The random effects will be used to estimate a common trend, property type trends, location value, and property random effects for properties that sold more than once. The machine learning algorithm will fit the observed characteristics (features) in a complex non-linear fashion. The model is applied to a small sample of 2,652 transaction in Phoenix (AZ) between 2001–2021. We only observe a limited number of property characteristics. The average out-of-sample MAPE is below 11%, which is as good or even better compared to the average appraisal error found in literature. The out-of-sample MAPE is even 9% for properties that sold more than once in the training set. In addition, our model provides indexes and locational heatmaps. These have their own uses and cannot be obtained with standard machine learning algorithms