How visual data can improve Automatic Valuation Models in Australia

By QUT's Associate Professor Viet-Ngu Hoang, Dr. Kien Thanh Nguyen and Dr. Andrea Blake

New research has investigated how incorporating visual data into Automatic Valuation Models to achieve a more meaningful and accurate estimate of value that is reliable for risk assessment and decision-making purposes.

Read the research here.

Valuation accuracy is fundamental to the efficient operation of the property industry including assessment of risk and decision making, projection of investment return and asset and portfolio management. Residential valuations are generally undertaken using the market approach.

This method relies on a comparison of the subject property to similar properties transacted in the market. Fundamental to this approach is the valuers’ ability to consider the characteristics and aesthetics, or visual desirability of a property.  

Recent history has seen the growth of property technology to streamline property operations. Like other property technology solutions, Automatic Valuation Models (AVMs) provide cost-effective estimates of value. AVMs are analytics models that use data from a variety of sources to provide an estimate of value.

Currently used AVMs that are available through commercial platforms, do not incorporate the visual desirability of a property in determining a broad range and frequently inaccurate estimate of value. The fundamental shortcoming of these AVMs is in not incorporating visual data into the estimate. Scholarly research has shown that models incorporating visual indicators deliver a significantly greater level of accuracy in determining an estimate of value. 

With the support of the Australian Property Research and Education Fund (APREF), a team of researchers in the field of property valuation, economics, and Artificial Intelligence at the Queensland University of Technology (QUT) have undertaken a research project with the aim of incorporating visual data into an AVM to achieve far greater accuracy. The outcome of this research is a more meaningful and accurate estimate of value that is reliable for risk assessment and decision-making purposes. 

Researchers developed an analytical framework using publicly available visual big data to capture visual aspects of residential properties. They built Convolutional Neural Network (CNN) models, in the ResNet architecture family, with street-view and aerial image data from Microsoft Bing and Google.

These CNN models are designed to construct a visual desirability indicator for each property. Importantly, this visual desirability indicator can be used in predictive analytics to capture the intangible qualities of a property such as attractiveness of street frontage, community green space, overall street aesthetics, and other attributes of urban design. Note that these attributes are not considered in existing commercial AVM solutions in Australia. 

In the pilot study, the research team applied the framework to three single-year data sets (2018, 2019 and 2020) covering housing located across 128 suburbs in Brisbane. They experimented many fusion models with differing visual desirability indicators constructed by the CNN models using either street-view images, aerial images, or both.

These fusion models are an extension of the commonly used a semi-log Hedonic Pricing (HP) model. Their empirical results show that the inclusion of the visual desirability indicator constructed the CNN model from visual image data has improved the accuracy of price prediction significantly.

The fusion model using street-view images has smaller root mean square of errors (RMSE) of 11 – 21 per cent, in comparison with the RMSE of the baseline HP models that do not have the visual desirability indicator. The fusion model using aerial images would perform even better, reducing the prediction errors by 20 – 29 per cent, varying from year to year. 

The results from the pilot study in Brisbane exposes the shortcoming of the current AVMs in achieving accurate estimates. It was the visual data, from publicly available sources such as Google or Microsoft Bing that had the capability of significantly improving the accuracy of valuation estimates.  

Undoubtedly, achieving greater accuracy in AVMs is a desirable outcome for many stakeholders in the property market. Whilst this research lays the foundation to achieving greater accuracy in valuation estimates by using big visual data, it is recommended that further research should consider the use of other visual data sources belonging to each specific property such as property photos, videos, or floor plans to provide even more accurate automated valuation. 

Learn more about the research here.

Contributors
All posts
Advertisement