Application of Geographically Weighted Regression for Mass Valuation using the Indonesian Land Agency Dataset for Bekasi, Indonesia

Author: Rahmat Sihombing

Sihombing, Rahmat, 2019 Application of Geographically Weighted Regression for Mass Valuation using the Indonesian Land Agency Dataset for Bekasi, Indonesia, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.

Abstract

Valid property transaction data in Indonesia is scarce because parties involved in a property transaction often report a false, lower transaction price to reduce their transaction tax liability. This has meant that in every mass valuation project administered by the National Land Agency of Indonesia (BPN RI), the sample size has never been sufficient to allow the currently employed Zonation Method to provide a complete prediction of land values across a city. A new mass valuation method that is fit for purpose when applied to a BPN RI-dataset is required. An extensive literature review was conducted to compare mass valuation techniques used worldwide, and Geographically Weighted Regression (GWR) was identified as the best potential candidate to replace the Zonation Method.

The performance of GWR modelling was tested on a typical BPN RI-dataset from the city of Bekasi in western Java. A road network dataset was required to generate data for ten of the 12 parameters listed in the current Mass Valuation Standards of BPN RI. A road network dataset had to be derived from the Land Parcel Map of Bekasi for this research because existing road networks from other sources had severe mismatches with the Land Parcel Map. Deriving a road network dataset from the Land Parcel Map was very time consuming because of the huge number of drawing errors in the Land Parcel Map that had to be corrected.

In the Bekasi case study, the GWR model had a mean absolute percentage error (MAPE) of 19.40 per cent, which was lower than the currently employed Zonation Method with a MAPE value of 10.80 per cent. Nevertheless, the GWR model solved the main problem of the Zonation Method; i.e. its inability to provide verifiable predictions for zones with fewer than three samples. Moreover, the MAPE value of 19.40 from the GWR model was well below the cut-off value of 30 per cent accuracy currently used by BPN RI.

The performance of the GWR model at non-sampled locations was estimated by out-of-sample estimation using Monte Carlo Cross Validation. The distribution of average percentage residuals from out-of-sample estimation resembles the distribution of percentage residuals from the in-sample GWR model. The correlation coefficient of the two distributions was 0.987. These two facts indicate that the GWR model does not have an issue of overfitting, and therefore it is very likely to maintain its prediction accuracy when predicting the non-sampled locations.

The main issue discovered when applying the GWR model was that a small proportion (7.51 per cent) of predictions at sampled locations had residuals greater than 50 per cent of the actual value. In the absence of overfitting, a similar proportion of predictions at the non-sampled locations are also likely to be inaccurate. Value zones were employed to detect potentially inaccurate predictions because predicted land prices in one value zone can be expected to be similar to one another. Local Moran’s I test and the coefficient of variation were employed to detect anomalous individual predictions in each value zone.

The problem of a lack of valid transaction data for mass appraisal modeling, while a big issue in Indonesia, is also a major problem in many other countries in the world. The approach taken in this study can potentially be adapted and amended in many other countries. The method is useful to provide accurate predictions at non-sampled areas. The key issue in applying the approach is the need for an accurate digital road network map.

Keywords: mass valuation, geographically weighted regression, indonesia

Subject: Environmental Science thesis

Thesis type: Doctor of Philosophy
Completed: 2019
School: College of Science and Engineering
Supervisor: Professor Andrew Millington