While stock and bond markets are receiving a lot of attention in the media at the moment, it’s important to not lose sight of the impact of Covid-19 on real estate markets and the influence this can have on people’s lives.

Many obviously own a house or apartment, and a substantial part of pension funds are exposed to real estate as an asset class. It is therefore extremely important to provide accurate valuations in times of increased stress and volatility.

Capital Economics’ Chief Property Economist Andrew Burrell estimates that COVID-19 could impact commercial real estate valuations by at least 8% in 2020. Quick and accurate valuations are therefore essential at this time.

Currently, the task of valuing real estate is predominantly performed by basing these valuations on similar buildings in the area, known within the industry as ‘comparables’.

The recent lockdown of much of the world due to COVID-19 has significantly reduced market sales, which has in turn made it harder for valuers to find comparable buildings sold in the area. The associated increase in uncertainty results in greater volatility making transaction prices less informative, which in turn affects volatility further, creating a feedback loop that makes it more complex to value property.

This is why, in part, certain UK property funds have now banned withdrawals. This has been implemented to protect their investors from making unfavourable trades, however this should not be happening for such an important asset class as real estate.

With the rapid advance in technologies applied to real estate, platforms have emerged which predict real estate prices using big data, a process called Automated Valuation Modelling (AVM).
Bart Melman – RE5Q

With the rapid advance in technologies applied to real estate, platforms have emerged which predict real estate prices using big data, a process called Automated Valuation Modelling (AVM). Using these techniques, it is possible to assist valuers with computer-driven valuations to make better and more well-informed decisions.

In this article, we show how automated real estate valuations from big data can help valuers in times of crisis, while also demonstrating how machine learning can improve the performance of existing valuation methods.

In times like these, the real estate market performs completely differently due to a lack of market activity and the resulting higher volatility. The spread in real estate prices increases and average prices can vary more quickly over time.

Valuing real estate in the same way as one would in normal market conditions therefore becomes more complicated. Before, one could compare two buildings in the same street or block, but now the comparative isn’t quite as obvious, or clear.

Suddenly, a valuation must consider how one neighbourhood relates to another, which is no easy task. Additionally, an increase of uncertainty makes the transaction prices less informative which means more observations are needed to build a clearer picture and counteract higher uncertainty at exactly the time when transaction volumes are falling.

Local trends need to be captured by the valuation model.

In the Global Financial Crisis of 2008 and subsequent crash in property markets, people discovered that prices became locally correlated. For example at that time in the UK market, certain types of building, or certain towns and areas were affected more than others. If three people in the same street put their house for sale, then prices are going to drop regardless of the rest of the market. These local trends need to be captured by the valuation model.

An ideal automated valuation model based on comparatives should make a trade-off between geographic distance, similarity and time of last transaction.

“Everything should be made as simple as possible, but no simpler.”
– Albert Einstein

The ideal valuation model should be flexible enough to capture the market, but not too complicated at the same time. As Einstein put it: “Everything should be made as simple as possible, but no simpler.” Furthermore, the predicted price should incorporate local prices and all data available to make a balanced prediction.

Existing models have mainly focussed on predicting real estate with building attributes, such as floor area, office quality and proximity to public transport. Even though it is simple, it does not incorporate prices of neighbouring buildings.

In addition, some approaches are incapable of incorporating non-linear patterns – an example is how the price per area of office space decreases for larger office spaces.

The approach which most closely matches that of a traditional valuation is to take the average price of all properties in the neighbourhood, which is called K-nearest-neighbours (KNN). Surprisingly, it is the only approach which directly uses the transaction prices of neighbouring properties.

An enhanced version is to take the average prices of similar building attributes, such as floor area, but these rules must be set manually and are not directly learned from the data. So as a valuation method, it has its pitfalls.

There is another way, which we call ‘AI Powered KNN’, which improves on the previous model by learning the similarity of buildings directly from the data.

The predicted price is a weighted average of neighbouring sales, where a larger weight is assigned to properties which are closer to each other, display greater similarity, and/or were sold more recently. As a result, the model learns to make the optimal trade-off between geographical distance and similarity using state-of-the-art machine learning.

For real estate investors, this valuation model provides interpretable results, which can be presented to all stakeholders in a clear and concise manner. For every valuation, it is evident which buildings are used to predict the price and what weight is assigned to each transaction price.

Investors can visually inspect weightings and can manually adjust them to reflect any additional information that has not necessarily been processed by the model. This model can also be used to find comparable investments, using similar weightings to recommend alternative properties.

COVID-19 is testing all aspects of life – from peoples’ living and working habits, to the response of healthcare, manufacturing and retail, the role of governments and supra-governmental agencies, and to markets and economics.

Crises like this also test new models and approaches, and are quick to show where older legacy systems and approaches need replacing. Real estate valuation is one area where this is happening.

AI-based approaches to automated valuation models over time can further be enhanced by factoring in both traditional and non-traditional sources of data.

Existing automated valuation models are too simple and fail to jointly incorporate a building’s attributes, local prices of neighbouring buildings, and non-linear behaviour – all of which are required to drive an accurate valuation.

AI-based approaches to automated valuation models over time can further be enhanced by factoring in both traditional and non-traditional sources of data. This is much needed at a time of increased market stress.

There is an exciting road ahead for innovation in the real estate sector that can help overcome some of the challenges felt by investors in a bear real estate market. A new way of doing valuations is just the tip of the iceberg.

Written by Bart Melman

RE5Q