Valuing property is an inexact and antiquated art, coming down as much to the gut feeling of the valuer as anything else.
Teun van den Dries thinks he can change that, using artificial intelligence to get an accurate price for every part of the world’s $200tn property market.
“We want to be the Google of the built environment,” says the softly-spoken Dutchman whose measured delivery belies his outsized ambition. “That is definitely the scale we are thinking of. We want to understand the value of every building in the world.“
Some big names are backing his vision. His Delft-based GeoPhy in January raised $33m in a Series-B financing round led by Index Ventures, to take the approach – already in use with clients in the UK, US and Netherlands – global.
Real estate — both commercial and residential — is the world’s largest asset class, valued at somewhere above $200tn, according Savills, the real estate agent. It is actually hard to be much more exact than that because valuing property is notoriously imprecise.
Subscribe to the Sifted Newsletter
“There is a reason it is described as more of an art than a science,” says Dan Hughes, chief executive of Liquid REI, a consultancy that advises property companies on the use of technology. “The valuation of a property depends on many factors, such as transport links and even the design of the building. These are difficult to measure so the valuer usually has to go to physically view the property. A lot of the time it comes down to gut feel.”
In the end, a property is worth what someone is willing to pay for it. The median difference between the actual sale price and the valuation is around 15%, says Van den Dries. GeoPhy’s AI-driven approach, he says, can bring that down to 5%.
GeoPhy, which employs 100 staff, is not the only company applying data to real estate. Hometrack in the US provides banks with automated valuations of residential property, and internet-based companies such as Rightmove have made residential property data much more widely available to everyone.
The commercial market — GeoPhy’s primary focus — is less well served, although Skyline AI, an New York-headquartered startup offers similar tools. Skyline raised an $18m funding round in July in a funding round led by Sequoia Capital and TLV Partners. New York-based Cherre is also applying deep learning to real estate data.
“In the last few years there has been a shift in how open the real estate industry is to looking at new technologies.”
“In the last few years there has been a shift in how open the real estate industry is to looking at new technologies like these,” says Hughes. “It is still at the early stages of quite a long journey, but in five to 10 years I think we could see the way we value buildings dramatically change.”
The idea for GeoPhy was born four years ago, when Van den Dries was buying a house.
“I needed to get the house appraised to get a mortgage. The appraiser actually rang me up to ask what the value should be. For me as a consumer that was helpful – I could get a cheaper mortgage, but it seemed a crazy way to do it.”
Together with his cofounder Sander Mulders, another former architect, Van den Dries began collecting data on the 8.6 million properties in the Netherlands, from places like the Land Registry, and combining them with hundreds of other data sources including recorded transactions, population density, crime rates and satellite imagery to create a valuation model.
“There are a number of factors that influence why an office in central London might be worth more than one in, say, Southampton. Population density, spending patterns, transport. We can model all of that.”
The company started by helping Dutch banks, still recovering from the 2008 housing crash, value the properties they had on their books, then expanded to mapping the UK housing market. Next came a call from Fannie Mae.
Van den Dries was partly bluffing when he agreed to help the US mortgage lender value its vast real estate portfolio – he had yet to hire the data scientists for the project, and it was a task that many larger companies had tried and failed to crack.
“One of the company’s rules was never to hire anyone from the property sector.”
Van den Dries was confident, however, that his tech-led approach would work. “We approached it with a fresh pair of eyes.” In fact one of the company’s rules was never to hire anyone from the property sector. “In the beginning it used to be a disqualifier if you had experience in real estate,” he says.
Now Van den Dries wants to extend GeoPhy’s reach to the 50 developed real estate markets around the world where there are plentiful public records on property, such land registries and ordnance surveys. After that, the challenge will be to create a model for countries where few records exist at all.
Still under construction
Apart from convincing a famously slow-moving industry to adopt its technology, there is some development work still to be done on the technology. Van den Dries admits the AI model doesn’t work for every situation. Unusual buildings, for example, are hard to price. “It wouldn’t work for something like the Shard,” he says.
It also doesn’t work particularly well for retail properties, in part because the retail sector is undergoing such rapid change, with online retailers decimating the trade of physical stores.
The number of property surveyors may decline [there are currently some 82,000 employed in the US] but not all of them will be out of work, says Hughes. “If I was buying a big office building for £100m I would want someone to hold my hand and advise me. But if you are valuing a portfolio of hundreds of companies you would want to automate that.”
Apart from anything else, an AI-led approach is quicker. Fannie Mae has gone from valuing its portfolio every three years to doing it monthly.
“We are actually kind of looking forward to the next property crash. It will be a good opportunity to test the model.”
A big question, however, is whether the model would cope with a property market crash. GeoPhy is trying to ensure it does – one of the company’s research projects is to see what data would have predicted the 2008 housing crash. Van den Dries says there were clear clues in consumer credit defaults, for example. “There are a lot of things people stop paying for before they stop paying their mortgage. We can track those,” he says.
But then again, each market crash looks slightly different from the last, and it may not be possible to prepare for the “unknown unknowns”.
“We are actually kind of looking forward to the next property crash,” says Van den Dries with a laugh. “It will be a good opportunity to test the model.”
Subscribe to the Sifted Newsletter