In a recent article I presented a mortgage application analysis tool that uses AI to validate applications. After reading that article, a client we worked with last year (Vin Vomero, founder of Foxy AI) reached out to see if I could write up a similar article on the work we did for them on real estate valuation.
The problem we worked on last year was a surprising one: Home price estimates are bad. Listings are being overvalued or undervalued all the time, and the valuation of a home is a key factor not only for the buyer and seller, but also for the underwriter on the mortgage loan for which the home is collateral. Automated Valuation Models (AVMs) are often used by financial institutions to make decisions on everything from home equity loans to credit card limits. It’s not only about home sales. Understanding the value of a home is critical to loss mitigation and credit risk management.
We live in a world of fast-paced technological change, and adoption of Artificial Intelligence is no longer an option. Rather, it’s a competitive necessity. Soon, valuation models without some level of AI will be obsolete.
To quote Rob May, CEO and Co-Founder of Talla, "It amazes me that some companies are slow to adopt AI because they aren’t sure where to apply it or how well it will work. I think they don’t understand that, by the time these things get figured out, they will be way behind."
Zillow’s announcement that they have incorporated image analysis into their Zestimate highlights this point. AVMs applied around the world need to be a lot better, because they have been objectively terrible until now. The fact that AI can get a 15% improvement in price prediction is an indication of just how poorly the traditional models were performing until now.
Why are home valuations bad? They rely on old data, numerical data, census data, IRS data, and state/regional sales data that may not be reflective of the investment put into the home. Even tying all these data sources together sometimes gives an incomplete picture. Although traditional AVM models are quick and cheap, they can’t truly take into account the quality and condition of a property without really looking at the property, thereby limiting the accuracy of a key tool in the financial services industry.
Why is this a hard problem? Well, traditional computer algorithms like to process numbers using formulas, not pictures of homes. There is no field on a listing for "one shingle on the roof is missing." To get this right, the AI model has to look at pictures and understand how the pictures affect price. More specifically, the model needs to learn about what room it is looking at, understand textures, colors, and indoor/outdoor objects. Think of this solution like a dating website innovation where until now, no one could see the profile pictures of the people, and instead they could only see their height, weight, and other biomedical data. Clearly, a picture is worth a thousand words.
This problem of integrating image data into home valuations is a really good fit for an AI solution. Even when implemented as an AI solution, getting one regression model to understand so much data about a listing, and getting it to understand all the special cases and exceptions, is a bit of a nightmare. I’m going to skip over the technical solution details of how crazy hard it is to built an AVM, and instead focus on the novel use cases that Foxy AI brings to the table.
The pictures on a listing can tell us a story about the property valuation that can be augmented by the traditional numerical, categorical, and text data associated with a real estate listing. Granite counter-tops, for example, indicate that money was put into the house. Subtle factors like the condition of the paint on the walls, tell us a lot about the expected sale price for a listing. The dataset used to train an AVM is basically the normal type of dataset that traditional AVMs are based on, but augmented with lots of extra data from pictures and other sources.
We delivered an initial prototype last year. That prototype we developed for Foxy AI has now grown much more mature, and has been trained on a really breathtaking dataset. Now, thanks to advancements in algorithms and computational power, we can take advantage of tools in a way that was never before possible. Understanding the gradient in quality across millions of properties is now a reality.
Foxy AI research has now brought to market this novel AVM approach, combining computer vision and Deep Learning, to assess the quality and condition of residential real estate. The whole system is exposed through APIs so that you can snap it on top of the other stuff you are using. The product uses advanced training techniques and a suite of in-house prediction models to classify what type of rooms make up the provided image set and determine the quality and condition of the subject property, to improve valuation accuracy.
There are really interesting approaches exposed by the API, like an endpoint to convert images into vectors, a way to find comparable properties (agents love that stuff), and a system for detecting objects, finishes, scenes and other stuff. And so you can decide to use the API to extract features from an existing listing, or just use the API to get better valuation data.
If you are relying on traditional AVMs, and you have been on the receiving end of inaccurate valuations, this could be really interesting for you to evaluate as a technology solution. Literally billions of dollars are at stake, and so even small gains in accuracy really affect Key Performance Indicators (KPIs) for reducing lending risk, increasing lending power, and so on. If you want to experience the magic yourself, sign up for the private beta.
Here is a link to Foxy on medium, and I hope you get a sense of why it is so cool. This technology opens up new possibilities like simulating what an imaginary house would sell for, and using that to do a knockdown and build estimate as a contractor. There are pretty nifty calculator capabilities that simply didn’t exist until now. The solution is all wrapped in a simple API: https://www.foxyai.com/doc-api/#api-endpoints
And so, in conclusion, AVMs are being disrupted, and financial institutions should take note.
If you liked this article on AI for real estate valuation, then press the follow button, clap the clap thing, and have a look at some of my most read past articles, like "How to Price an AI Project" and "How to Hire an AI Consultant." Also, check out Foxy AI. In addition to business-related articles, I also have prepared articles on other issues faced by companies looking to adopt deep Machine Learning, like "Machine learning without cloud or APIs."
In an upcoming article, I will present something we have been working on for quite a while, that helps enterprises to automate their analysis of unstructured reports during internal audits.
Until next time!
-Daniel
[email protected] ← Say hi. Lemay.ai 1(855)LEMAY-AI
Other articles you may enjoy:
