| Summary: | Mass appraisal of residential real estate is desired and often required for asset valuation,
property tax and insurance estimation, sales transactions and urban planning.
Multivariate linear regression models, referred to as hedonic pricing functions, have
been used to 'unbundle' the characteristics of a dwelling by expressing its price as a
function of its mix of attributes. However, the relation between the value of a dwelling
and its intrinsic and extrinsic characteristics is complex and generally nonlinear. Consequently,
this study attempts to capture this inherently complex relation through the
use of Artificial Neural Network (ANN) models and investigates their ability to predict
residential real estate values compared to traditional hedonic techniques.
Researchers in the real estate appraisal industry have recently used ANNs to overcome
methodological restrictions such as nonlinearity and noise that result from the
use of multivariate linear regression techniques. Detailed locational factors, however,
have failed to be adequately represented in their models. In my work I extend current
research efforts by explicitly incorporating 'space' into ANN models. Through integrating
ANN techniques and Geographic Information Systems (GIS), the extraction, transfer
and recognition of spatial attributes—such as average family income or secondary
school provincial examination performance—can be facilitated. Results indicate that
ANN models outperform traditional hedonic models. Further, the inclusion of locational
attributes significantly improves the ability of both models to predict the value of a
dwelling. === Arts, Faculty of === Geography, Department of === Graduate
|
|---|
