A comparison of algorithms and techniques used in automated valuation models: Decision support for residential property appraisals.

摘要

This research compares algorithms and techniques used in automated valuation models. Seven different automated valuation models (AVMs) were compared and analyzed using property from Southern Orange County California. Two additional models were derived from the neural network based AVMs. These hybrid models utilized a neural network to select the comparable properties and then a simple mathematical calculation was used to derive the value estimate for each subject property. The two neural network methodologies employed were a three-layered feed forward network and a re-configurable three-layered feed forward network. The two neural networks differed in that the re-configurable neural network pruned any connection deemed unnecessary in finding the optima. One AVM was constructed based on multiple regression analysis. Another AVM used was Experian's Valuepoint. Results from Valuepoint were obtained from a third party intermediary. The AVMs were compared based on MAPE, RMSE, number of estimates within a 10% threshold, and best overall rank. The Hybrid Model based on the re-configurable neural network was able to outperform the other models tested. Ten cross validation trials were conducted and the results strengthened the results generated by the Hybrid Model.