Academic journal article The Journal of Real Estate Research

Predicting Housing Value: A Comparison of Multiple Regression Analysis and Artificial Neural Networks

Academic journal article The Journal of Real Estate Research

Predicting Housing Value: A Comparison of Multiple Regression Analysis and Artificial Neural Networks

Article excerpt

Abstract

This article compares the predictive performance of artificial neural networks (ANN) and multiple regression analysis (MRA) for single family housing sales. Multiple comparisons are made between the two data models in which the data sample size, the functional specification and the temporal prediction are varied. ANN performs better than MRA when a moderate to large data sample size is used. For the application, this moderate to large data sample size varied from 13% to 39% of the total data sample (506 to 1,506 observations out of 3,906 total observations). The results give a plausible explanation why previous papers have obtained varied results when comparing MRA and ANN predictive performance for housing values.

Introduction

This study compares the predictive performance of multiple regression analysis (MRA) and backpropagation feed forward artificial neural network (ANN) for single family residential property value. The two data models, MRA and ANN, are compared using different functional model specifications, sample (training) data and evaluation criteria. The same specification improvements benefits both the ANNs and the MRAs, and a plausible explanation is provided as to why other studies that have compared the MRA and ANN data modeling tools to predict the value of residential property have had varied results.

To fairly compare the data models, one should address the possible methodological problems associated with each data model that might distort its performance. The studies that compare the MRA and ANN data models are identified. In addition, several studies dealing with the MRA data model specification are also identified and applied to the data set used in this study. Inherent problems (in implementing the MRA and ANN data models) that may affect the performance of each model and hence affect the results of previous studies are discussed. These findings are then applied to a new data set for comparison of the two data models. The constraints and results of the comparison are then given along with conclusions.

Although the standard feedforward neural network with backpropagation learning is used for the comparison in this article, experiments were conducted with multiple learning variations such as enhanced backpropagation, backpropagation with weight decay, quickpropagation, resilient backpropagation, backpercolation and counterpropagation. Various ANN architectures such as ARTMAP, GAUSSIAN, regression neural, etc. were also examined. After hundreds of experiments and multitudes of architectures, the standard backpropagation was found to perform better than the other neural network architectures examined for this application. In addition, since several other studies compared standard backpropagation neural network with MRA with varied results, a change of the ANN model would not have allowed a direct comparison between the results and those of previous studies. Thus, the question and focus of this study is why have some studies concluded MRA is better while others have concluded that standard backpropagation ANN is better for predicting sold property value.

Implementation Issues

Some studies have demonstrated the superiority of ANN over MRA in predicting housing values (Tsukuda and Baba, 1990; Do and Grudnitski, 1992; Tay and Ho, 1991/1992; and Huang, Dorsey and Boose, 1994;). Other studies (Allen and Zumwalt, 1994; and Worzala, Lenk and Silva, 1995), however, have noted that ANNs are not necessarily superior. Because of an ANN's ability to learn and to recognize complicated patterns without being programmed with preconceived rules, it can easily be applied with little knowledge (statistically) of the data set. Unlike regression analysis, an ANN does not need a predetermined functional form based on the determinants. This feature of an ANN is important, since several studies (Grether and Mieskowski, 1974; Jones, Ferri and McGee, 1981; and Do and Grudnitski, 1993) found age has a nonlinear relationship with housing value (for the data set used in their study). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.