Data Driven Modelling using Machine Learning.
The Challenge
Since the 1970s, predicting the impact of urban developments, such as new transport infrastructure, on property prices has been done using a standard modelling technique called the Hedonic Price Method (HPM). As price is a measure of “willingness-to-pay”, i.e., value to the purchaser, it is challenging to model, and HPM models have several weaknesses. These include typically using linear regression and making it difficult to effectively represent spatial and temporal information. Many machine learning (ML) algorithms have been developed to address the limitations of linear regression. However, such algorithms are not as established as HPM approaches. This project aimed to explore the comparative performance of HPM and several ML algorithms in terms of the predictive accuracy and model transparency with different representations of the data, including the use of spatial and temporal information.
Partners
The project partners were the Urban Development Institute of Australia (NSW) and the University of New South Wales (UNSW).
The Solution
The research focussed on the proposed North-South Rail line in Western Sydney. Technical issues prevented the comparison of HPM and ML approaches for the dataset. Therefore, the main objective of the research was to implement and evaluate a machine learning predictive modelling and its application to problems of value prediction in an urban development scenario.
The first stage of the project was to train a predictive model for dwelling prices based on a dataset of historical property transactions (actual real-estate sales) for Sydney, which were integrated with geographical and demographic features, such as distances to points-of-interest for each dwelling and employment figures for local areas. A range of machine learning algorithms were applied and their predictive performance evaluated. The second stage of the research was the simulation of the future city using a scenario-based approach, by varying the parameters of the design process. In each scenario, the model trained in the first stage was applied to the synthetic city to predict value for every dwelling. Results show that although transport-related developments represent only a minority of the envisaged future city, they do generate value uplift and contribute to a clear reduction in urban sprawl.
Impact
The ability for government to explore future land use changes scenarios as driven through property prices and land availability via a simple visualisation interface has been trialled through the UrbanAI project. The modelling has provided insights to both government and industry on the pros and cons of siting new train stations and subsequent urban development connecting the new airport in Western Sydney to other key employment centres across Sydney’s Western Parkland city.
Traditional valuation is a skilled and time-intensive activity, with human valuers employing a range of methods, from physical inspection of a property to database and online searches. As the value of property as an asset class increases, so does its economic importance. The need to be able to rapidly achieve accurate and up-to-date valuations has led to increasing use of automated valuation models. However, as with most critical applications of statistical and machine learning algorithms, a valuation which is based on an automated valuation model will still pass through a number of stages, often including expert-coded rules, and human oversight. This project could be extended in several ways. For example, price prediction accuracy could be improved.
Contact
To learn more, contact FrontierSI at contact@frontiersi.com.au.