Alongside his work as a professor at ABS, Francke works part-time at Ortec Finance, a provider of technology and solutions for risk and return management. He obtained a PhD in econometrics from Vrije Universiteit Amsterdam, completing a thesis on the development of residential property valuation models. 'It's still really interesting to use econometric models to capture a framework that gives the value of a property. Explanatory variables are things like year of construction, floor area, etc. It’s never simple. There are always interactions between the variables and value can't be measured linearly. Location also plays a role: central Amsterdam is more expensive than the suburbs.'

Quickly and widely available

Machine learning models are on the rise when it comes to measuring property values. 'These models have been used more and more often in the past 5 to 10 years but the academic world had already done so earlier. What we’ve seen up till now is that machine learning models perform just as well as econometric models. Speed is a major plus. Research that once required me to write an entire thesis can now be done with machine learning in a single afternoon. You put the available data in an easy-to-access model and you get an estimated value in return. Because machine learning is so freely available to anyone, we’re witnessing the democratisation of the valuation process. And in this sense, I’m very enthusiastic about machine learning.'

At the same time, there are factors that sometimes act as impediments. 'One is an inability to explain. If you want to determine the value of a home under the Wet Waardering Onroerende Zaken (Property Valuation Act), you should be able to explain how you arrived at that value. But then you run into the ‘black box’ of machine learning. How the calculations are made is not really visible.' Francke also encounters this limitation in his work for Ortec Finance. 'We embarked on machine-learning models together with TU Delft and major market players active in AI. The results from these models don’t always make sense at first glance. This is why, until now, we’ve been using machine learning mostly as an alternative model to verify the results derived from econometric models.'

More useful for homes than offices

So explicability is a problem. Another issue is the lack of a confidence interval. 'If a consumer takes out a mortgage for the full purchase price of a house, then the bank will need a very accurate valuation. You’ll then want a confidence interval at the level of the individual but this is not a standard output of machine-learning models. With econometric models, such intervals are an integral part of the methodology. They allow you to say, for instance, that there is a 90 per cent chance that the value of a given house will lie between amounts A and B.' On the plus side, the more recent models give very reliable estimates for large volumes of housing. 'It’s tricky for a single detached house but, for a property on an estate with hundreds of similar houses, machine learning is in fact very reliable.'

Despite the shortcomings listed above, machine learning will continue to grow. Francke cites the research done by 1 of his PhD students, who is using an image recognition model to arrive at accurate valuations based on hundreds of thousands of photos taken in and around residential properties. 'It’s interesting, partly because we don’t know yet what will come out of the study. Apart from image recognition, there are opportunities in text recognition. Advertising copy on Funda can already tell you a lot. This is also being looked at.' The professor thinks that, in the coming years, machine learning will be very significant for the liquid residential property market in particular. 'In the market for commercial property, such as offices, you’re dealing with information that’s in short supply and difficult to generalise. In this market, you benefit most from econometric models. Machine learning requires large quantities of data. For residential properties, especially those that are very common, there’s usually enough. So if you want to buy or sell a terraced house in, let’s say, Almere, then machine learning definitely serves a purpose.'

Francke believes that the application of machine learning will rise exponentially in many different facets of the residential property market. 'I was recently at a presentation by a German buyer of such properties, which are structured into portfolios and placed with institutional investors. They had developed an app to determine the dimensions and condition of a property on the basis of photos. Machine learning may well suffer from an inability to explain and a lack of individual confidence intervals but there are plenty of applications where these are not major issues.'

A side effect of the advent of machine learning is that companies can no longer differentiate themselves simply because they have econometric models. 'Nowadays, any student with some basic knowledge of Python can call on all kinds of models. There’s no differentiation now ‒ technology is a democratiser. But the models do need good data and any differentiation will therefore chiefly be in this area. Access to good data will determine a company’s success even more than access to technology.'