AI in Property Valuation: Exploring the Future of AVMs with Anders Lund Francke #3

In this episode of Mortgages Covered, Richard Kemmish, EMF-ECBC & ISMMA Consultant, brings together Anders Lund Francke of Eiendomsverdi AS to explore the use of Artificial Intelligence in property valuation.

This episode of Mortgages Covered explores:

•  The growing interest in and use of automated valuation systems (AVMs) and advances in statistical modelling for residential property valuation.
• The core principles underpinning robust valuation models, including transparency, explainability, and confidence management. 
• The management of valuation uncertainty and its role in risk management and bank competitiveness.
• The complementary roles of AVMs and physical valuers.
• The emergence of “nowcasting” techniques to increase accuracy and market responsiveness.

Transcript

SPEAKER_00 0:00

AI is a topic that we all use every day, whether we are cheating on our homework or preparing for a podcast for the EMF ECBC. But today we're going to look at it when it comes to more important things like valuing properties that underline so much of the banking sector. Welcome to the EMF ECBC podcast. I'm delighted to be joined here in Stavanger today by Anders Lundfrank from Eindomsverdi, which is a leading research company working on the valuation of real estate here in Norway. Anders, thanks very much for joining us today. Thank you for having me. Tell us a little bit about how you currently use AI or machine learning to value properties.

SPEAKER_01 0:46

AI is a new kind of what we define. How do we define AI? According to how EU defines AI, AVMs are AI. It is an automated system, so that's nothing new for us. As of the new, I mean the development is going quite fast, both in the LLMs such as ChatGPT, Gemini, etc., but also more sophisticated statistical models such as random forest, XG boosting, and those kind of techniques are has a huge potential that we that we all have seen. We are very much looking into especially the statistical part of the of the modeling in the in the actual valuation of the properties. We all obviously also use kind of the ChatGPT in our daily work. But what is very important for us when we utilize models, regardless, is some principles that we work after. We need our models to be explainable, we need to be transparent. We have to tell a bank how did we actually compute this valuation if they ask. For instance, we every week when when our team gets to work, we have the actual sale prices on the properties that were sold last week and we compare it with the model. And then we compare that with that it's unbiased, as I said, on average we should be on the mark, even in changing market conditions. So we if a bank reports LTVs to the market and they they update with our models, which they do, you can trust that LTV. Um we also measure confidence. So if if we say we're uncertain, we well, other way around, if we have high confidence, we should be confident as well. So kind of the distribution should be very narrow and we should document that. So we share all the performance report regularly with our risk management in the banks. So the risk management can say, okay, because the banks your job is to take risk, right? Our job is to provide transparency, data that they can take enlightened decisions. So you could say that, okay, we say we we overestimate some properties by let's say more than 30%. Let's say it just that give it, take a number. Let's say it's 1% that we overestimate by more than 30%. Then the bank can say, okay, we're happy with, we could be happy with that risk if it's if it's automates a lot of processes, if it enables us to grant more mortgages, you have all the data there. Or you can say, no, we're not confident with that. We need to have higher haircuts. We don't want anybody in the above 30 threshold. Uh it's or you can say, well, in this and this area, we don't want to utilize C A VM. We should have a fiscal value or a second opinion on that. So all this kind of data from our side is the data and the information provider and the modeling and just enable the bank to take enlightened decisions.

SPEAKER_00 4:43

So I get the point about there's no bias, there's no systemic bias. On average, you get it right. But how how confident can you be about how inaccurate you are? So if you say it's a million and you say it's plus or minus 100,000 Nokia, is that a reasonable number? Does it vary for different valuations? Sometimes it'll be plus or minus 10. Huge variation, huge variation.

SPEAKER_01 5:06

So what we actually see is there are some pros and cons between, I mean, you have different sorts of valuation sources. You have typically AVM providers such as ourselves, or you have the typical physical value, which in Norway could be a real estate agent or a surveyor. Um what we see is that we know that we are 100% data-driven. So the biased part is a strong point for us. Uh, we see that from the physical perspective, they're more to the mark on having more narrow distributions, but they tend have the tendency to be a bit biased because they actually go into your home, you get a coffee, what do you need this valuation for? Well, I need it for mortgage purposes. Okay, then it's high if you if you're into a divorce situation, are you a buying or seller, etc.? So they're kind of not unbiased, it's a human invention. Um, what we see is that if the property is very standardized, we are very good uh also on the on the distribution. So if you look in apartments in the big cities, typically the building bodies done by the housing co-op, um the bathrooms are very similar. This the distribution and you have a lot of data. So you you kind of have a lot of sales that you can compare it with, then the distribution is super narrow. I mean, if you're in more rural districts, uh typically harder for us, then it's harder for the evaluator as well. Um what the AVM is our weak point is if the property has been refurnished or changed a lot since the previous sale. Because we we emphasize and we we see that it's a lot of information in the previous sale. But if you buy a property to refurnish it and you actually do that refurnishment, our estimate will typically be way too low. Um, then you actually need a surveyor to go in and check that this job has actually been done. Um so that's kind of the the main advantage or also a disadvantage for us is some properties we do not have a lot of information because it hasn't been in the market for many years. Well, then we then we're guessing how big it is, for instance. We're guessing how many floors the uh the building has. And obviously then it's you just adds a layer of uncertainty. And how you can manage that in a responsible way is that well, okay, you you tell the bank that this is high uncertainty, let's make a big haircut. So you can say that, well, we do a haircut to the point where we're 80% certain that it's worth at least this. And then you can and then you can combine that with LTVs. You can say that okay, if this you have done a haircut 80% certain, which is probably quite high. If the LTV is, let's say 30% or something, for um my parents, for instance. If they were to refinance to buy a car and they take the collateral in the home, it's fine. You don't have to ask them for a surveyor who typically will cost a thousand euros or in that range. So it's just it is part of the digitalization optimization, which in the end ends up also good for the customer. It enables also competitiveness between banks, um, taking down the transaction cost of changing banks.

SPEAKER_00 8:30

I can see when you've got a lot of uncertainty about the value from a risk management point of view for a bank, you okay, we just take a haircut. Um but for an estate agent wanting to put it on the market, that starts to become a bit emotive.

SPEAKER_01 8:43

Yeah, so so it's it's not the same model for the real estate agent. We do also have the real estate agents as clients. We have the data. So for a real estate agent, they prove we we get daily data on the transactions in the market. So they go into our system, they look into if I if I were to sell your property, I will show up well prepared. So I know that this property in your neighbor was sold two weeks ago for four million. And they know every every property is and they have a discussion with the seller. What are you willing to sell for? This is the probable probable market value based on these comparables, and they do a bespoke set of comparables themselves. Um so it's a different process for the real estate agent when they uh sell the property. And it's also they do that for for selling the property, um, but they are also involved in the the refinancing bit the example I mentioned earlier, or just that our model is off the mark, the bank can typically ask, well, um we have a local real estate agent who is very good at the market values. Could you please do a valuation of the property for us?

SPEAKER_00 9:54

That's what that's what I was going to ask that. Is is the ultimate fallback an old-fashioned valuation? If I just don't like what your model comes out with, I want a human being to come and check this who understands the local market. Yeah. And do you get that often?

SPEAKER_01 10:06

Yeah. It's it's it's a combination, and we've been in this market for many, many years. Um I think other jurisdictions in Europe, it's I feel I I don't know that I'm not in the other jurisdictions myself, but I feel that it's a it's a bit um valuators are afraid of AVMs.

SPEAKER_00 10:24

Yeah.

SPEAKER_01 10:25

Um but our impression is that the valuators in Norway um very busy calendars. It's very difficult to get by uh value. Uh so it's it's a combination. It's a combination of the of um R models and physical valuators, such as surveyors and the real estate agents. So they are running the streets. So yeah, um it's room for both uh actors.

SPEAKER_00 10:52

Well, lots of people are frightened about AI taking over their jobs. Yeah, me too. Well, we all are, yeah. You're at least on the right side of it.

SPEAKER_01 11:01

Yeah, yeah, it's uh it is a bit scary at times, right? It's uh some nights I wake lay a wake up and it's like uh I'm going to lose my job tomorrow.

SPEAKER_00 11:13

Um But it's a very real concern for realist for valuation agents, right? They spend years learning, getting all this experience. Is there opposition really? What else could I say?

SPEAKER_01 11:25

The physical valuation part is I think a large part of that is actually to do an inspection of the property. Right. So so I think you still can't get them AI today. Yeah, no physical inspection. No, exactly. And and you only have the data that you have available. So you need to get that from and that would be in the in a point of time. So unless I mean you could probably that private person can upload some data to to evaluate things. But um things are um I'm afraid to lose my job to AI as well, but my impression is that uh things are changing fast. It looks really fast, but in the end of the day, it takes a bit of time regardless. Um But yeah. I I I think I think the I think the value profession still will be there. You will be more efficient and more accurate using data and and tools. Uh but in some cases, many cases you still need uh actual value.

SPEAKER_00 12:28

Yeah, I guess that's the defense against AI, isn't it? It's not gonna make me redundant, it's gonna make me much better at my job.

SPEAKER_01 12:33

Yeah. Um that's my impression so far. And there are as you said, there are pitfall with You're just working in another way. Um, my lead data scientist, actually, uh in our team, is a tech guy, uh trained tech, I mean um IT uh guy. He showed up at work and said, I've written my final word of code. So he's is now only kind of writing in his natural language, but and he was like, Well, yeah, I'm afraid of my job now. But I think the profession is still kind of you need to overlook what you're actually doing, you need to understand what are we solving here? Uh, and that's a lot of pitfalls, and the AI is still making a lot of errors.

SPEAKER_00 13:22

I suppose the other big problem with AI is just persuading people that it's right. And when you talk about things like valuation for banks, thinking about things like bank capital, that's pretty important. Yep. How much do the regulator look at your model and how much do they hold it to the hold your feet to the fire and say, prove that this really works and provides a conservative and meaningful valuation?

SPEAKER_01 13:43

They are more more interested now than they used to be.

SPEAKER_00 13:48

Um we are not more reliance on it.

SPEAKER_01 13:51

Yeah, we we are not regulated by the Norwegian FSA, uh, but the banks are. So they are utilizing our model. So we we have a very good dialogue with the Norwegian FSA. In Denmark, for instance, the AVM is actually under under the Danish FSA, so that they are uh under stricter regulations than we are, but we are indirectly through the banks. And we yeah, we they are very interested, and as um we are they are very transparent about our accuracy and the reports and the data as I showed for the risk management, we share the same reports with uh with the Norwegian FSA.

SPEAKER_00 14:29

And is it gonna go the same way as the Danish? Is do you see all AVMs across Europe ultimately being regulated by the banking regulator? I have no idea.

SPEAKER_01 14:38

I'm not I'm not I'm not a lawyer, I'm not an expert in the regulatory space. Do you think it should? Uh there are some pros and cons uh for us. I mean from my from a nanos ready perspective. I think it will uh it would be more difficult to compete with us um because we have a very strong position, uh so it would kind of be a regulatory um uh protection. For me personally, it probably will be a bit more reporting, which I not necessarily like, but uh maybe you have AI for that. Well yeah. You can do the reporting by that.

SPEAKER_00 15:15

Yes. Um one final thing you I want to ask about. You clearly look you you're basing it on historic data. Yeah. And the one thing we know about house prices is they can change pretty rapidly over time, either up or down in the crisis.

SPEAKER_01 15:28

Yeah.

SPEAKER_00 15:28

How how much can you take that into account if the last property that was sold two years ago and you know that properties in everywhere else have gone up but by an unknown amount, can you really apply those changes? How can it be forward-looking?

SPEAKER_01 15:42

Uh it's not and it's a very important um thing for us that we are not forward-looking. We are now casting, which means that the model is only seeing what it sees. So you always, or there was during the financial crisis, which was uh perfect storm for the residential uh housing market in Norway, you saw that the volumes dropped by uh 30-40 percent, something like that, but you still have transactions going on. Same in during COVID, uh the transactions went through, which means that you have comparable sales. So you can adjust the market factors. And we we do measure this. Um, actually, on the during COVID, we had we had webinars weekly with risk uh departments in the banks where we showed them how well the model actually worked. And that that is also a strong suit for an AVM. It's impossible to do that for a physical evaluation, right? Because it's you just don't have the data, you you're not able to update that on a daily basis on a on a bigger scale. Um, so the model is actually working quite well to adapting markets, and we get that question a lot what if you have to sell in this bad market? That's not what the model is for, because you have to have a willing seller and a willing buyer, and the willing seller is equally important as a willing willing buyer. So we're not trying to predict the forced sale or if you have to, if you are in in a hurry, you have to sell in this bad market. That's not what we're doing, we're just observing, and that's that's it. So and we're very we try to be very clear on that. Uh it is a very good question we and we do get it a lot. Um but yeah.

SPEAKER_00 17:27

It's a very relevant point about a a willing buyer, a willing seller. If you're selling a standard flat in a residential part of Oslo, yeah, you'll find someone pretty quickly who wants to buy a flat there. Yeah, if you're buying an unusual property in a small town, who knows if the right person's gonna turn up. Yeah. And is that just part of that uncertainty that you spoke about earlier? Yeah, exactly.

SPEAKER_01 17:47

And what we actually see, we have done some research on that, uh, is that if you are a forced seller in uh illiquid market, the the discount will be bigger. Yeah. So typically if you if you're um yeah, the market is just not that liquid. You as you said, the unknown if the seller is no, sorry, the buyer is there, you typically have to go much further.

SPEAKER_00 18:09

But but that's a different number. So there's the expected value, the deviation around that, and then there's the forced value. Exactly.

SPEAKER_01 18:16

So we just see that that discount varies according to what and not just that it's um in rural areas, but also special properties, uh super expensive things that just doesn't have that thick market. It's it's um yeah, it's more it's more a discount from the predicted market value.

SPEAKER_00 18:38

And f final question before I let you go. Um Are the models getting better? Are the models gonna be better than they are in five years' time, say?

SPEAKER_01 18:45

Yes.

SPEAKER_00 18:46

We are improving every year.

SPEAKER_01 18:48

Uh so we have a team working very hard for that. Both we get more data every year. So the data is very important. It's um we say that one-third of the job is the data, at least, kind of preparing the data. You always everybody working with data that know that you get crap data, and that could it's important to prep those data before you do the modeling. Um then we use the modeling, and then it's the testing. So that's the three parts, and we test regularly, and we see that we improve gradually over time.

SPEAKER_00 19:23

Yeah. Thank you. Thank you very much for speaking to us today, Anders. It's been a fascinating area. I think it's an area that we're going to come back to in future editions of this vodka this podcast. But thanks for joining us, and thanks very much to everybody for listening. Thank you.