Digital real estate. When algorithms do appraisals.
For over 20 years now, real estate appraisals in Switzerland have been performed with the help of algorithms. Cutting-edge machine learning techniques can further improve the accuracy of current estimates. However, this comes at the cost of transparency and increases volatility. This article covers these issues and more concerning digital real estate.
Machine learning is becoming increasingly important
Machine learning is a subfield of artificial intelligence and its purpose is to predict outcomes based on data patterns recognized by an algorithm. This generally involves analyzing huge amounts of data. As it does so, the algorithm learns from existing data and can increase its accuracy.
Over the past decade, machine learning (ML) has become considerably more important in many areas of our lives. A key driver of that was the steep rise in the processing power of modern computers. That trend has also had an effect on real estate appraisal.
Hedonic pricing models as the standard for appraising properties
Machine learning is anything but new. More than 20 years ago, hedonic pricing models, which are also considered to be ML methods, became established as the standard when appraising residential properties. These models have been continuously improved since they were introduced, thus enabling these computer-based methods of valuation to be used more and more even for relatively small investment properties. The price for a piece of real estate is determined on the basis of various property characteristics. The algorithm optimizes its modeling in such a way that the valuation error is a small as possible. Hedonic models have improved the degree of transparency on the Swiss real estate market considerably.
Research focuses on new machine learning algorithms
The success of hedonic models and the improved availability of real estate data have given an added boost to real estate market research in Switzerland. However, the interest in research is shifting to new ML algorithms. Examples include ML techniques known as random forest and gradient boosting. The way they work is based on decision trees.
Those trees represent a visualization of successive, hierarchical decision-making criteria. In the case of real estate appraisal, this allows the program to determine and/or predict the price of a property. The general rule is that the more complex and comprehensive such a tree is, the more accurate the forecast will be for all the properties in a training data set. However, it is important that the model not be over-specified.
Complex modeling using neural networks
There are other ways of modeling real estate prices using ML techniques besides decision trees. Those methods often involve the use of artificial neural networks, which simulate the way the human nervous system works with its vast number of neural connections. The structure of the neural network allows the complex combination of and interaction between the individual explanatory variables.
Better results, but models remain a black box
The possibility of improving real estate valuation using state-of-the-art ML techniques or artificial neural networks also has its down sides. While hedonic models have the advantage of enabling interpretation, modern ML techniques are less transparent, which poses certain challenges to their use in daily practice – for instance, during client meetings.
Another disadvantage of cutting-edge ML techniques is the higher degree of volatility, caused by use over multiple quarters, than that of traditional hedonic models. The result is that the value of an individual property determined using these models fluctuates much more heavily from quarter to quarter than it does with hedonic models.
Property appraisal models still have a number of deficiencies
Modern ML techniques also have the same weaknesses as conventional hedonic methods. Unusual cases prove to be stumbling blocks for them. For example, luxury homes or those that come with special "character" are more difficult to appraise. Depending on the chosen model, it will involve a certain time lag. That is because the transaction data used for modeling is, at best, taken from the previous quarter. Therefore, none of the models can reflect up-to-date price changes on the market. This problem is likely to continue into the future.
Depending on the algorithm chosen, there will also be room for improvement especially in the data collected. After all, many algorithms today still need to be fed by hand. The microlocation – that is, the immediate vicinity surrounding a property – is automatically determined by most models based on the address entered, for example, but the quality of the finishings and the condition of the building need to be evaluated by a person. This is one area where the data entry process could be increasingly automated through machine learning.
Conclusion: Machine learning is helpful but is not a magic formula.
In conclusion, it can be said that modern-day machine learning techniques can improve the quality of predictions regarding property valuations to a certain extent. However, that comes at a cost: lower transparency and higher volatility over time. Consequently, state-of-the-art ML methods, such as those used in real estate financing by banks, are, for the most part, being used hesitantly for the time being. Hybrid approaches combining conventional hedonic models with cutting-edge ML techniques could serve as a door-opener in this area.