Artificial neural networks modelling for mass appraisal of properties

• Main Author: Yacim, Joseph Awoamim
• Other Authors: Boshoff, Douw G.B.
• Language: en
• Published: University of Pretoria 2018
• Subjects: Property valuation; Mass appraisal; Artificial intelligence; Neural networks; UCTD
• Online Access: http://hdl.handle.net/2263/64741; Yacim, JA 2017, Artificial neural networks modelling for mass appraisal of properties, PhD Thesis, University of Pretoria, Pretoria, viewed yymmdd <http://hdl.handle.net/2263/64741>

This thesis extends the use of artificial neural networks (ANNs) optimisation and training algorithms including the Powell-Beale conjugate gradient (PBCG), scaled conjugate gradient (SCG) and a hybrid system of particle swarm optimisation (PSO) with the traditional back propagation (BP) in mass appraisal as a first attempt. The goal is to verify the comparative performance of ANNs with the traditional hedonic regression and some other modelling techniques including geographically weighted regression (GWR), spatial error model (SEM), spatial lag model (SLM), additive nonparametric regression (ANR), M5P trees and the support vector machines (SVMs). The methodologies are applied to data of 3232 sales transaction of single-family dwellings sold during the period, January 2012 to May 2014 in Cape Town, South Africa. The analysis was done in categories such that the best performing method in each category is selected for a final comparative analysis. The results reveal that semi-log model, SEM, normalised polynomial kernel function support vector machines (NPKSVMs), ANR and the Levenberg-Marquardt trained artificial neural networks (LMANNs) performed best in their respective category. The study also demonstrates the practicability of building hybrid systems in mass appraisal, unfortunately, the hybrid models produces an unexpected results relative to the standalone ANN models. Furthermore, the five best performed models were subjected to three different tests namely, prediction accuracy within the 10 and 20%, model performance and reliability ranking order and lastly explicit explainability ranking order. The final results reveal the LMANNs to outperform the ANR, semi-log, SEM and SVMs in the first two tests, but when the explicit explainability ranking order test which consist of simplicity, consistency, transparency, locational and applicability within the mass appraisal environment was performed, the LMANNs failed the test. The results demonstrate the SEM as the most preferred technique because of its transparency, locational advantage and ease of application within the mass appraisal environment. Furthermore, it is inferred from the findings that having superior predictive power is imperative, but most importantly is whether the model can practically and effectively be used in mass appraisal of properties. The black box nature of the ANNs inhibits the production of sufficiently transparent estimates that appraisers could use to explain the process when required as a defence before a tribunal or in a formal court. This thesis contributes to knowledge as follows: i. Analyse the significance of spatial variation of property prices, with Cape Town, South Africa used as case study; ii. Build a hybrid system of PSO and BP in mass appraisal; iii. Improve the training of ANNs in mass appraisal with SCG and PBCG algorithms; and, iv. Extend the use of GWR, SEM, SLM, SVMs, ANR and log transformation of variables into the South African property market context. === Thesis (PhD)--University of Pretoria, 2018. === NRF CSUR13092648264, Grant No: 90311 === Construction Economics === PhD === Unrestricted