Know good | catch bad, with Sjoerd Slot

Criminals keep changing the game. Which is annoying, don't they know our models depend on them re-enacting the past behaviours over and over and over again? Sygno have flipped the script by modelling the good behaviour of your customers... because they're not trying to outmanoeuvre you. Also, there are a lot more of them. Sygno's automated machine learning technology generates monitoring models for Anti-Money Laundering and fraud to make transaction monitoring more effective at a lower cost.

You can start a conversation with Sjoerd over on LinkedIn: https://www.linkedin.com/in/sjoerdslot/

Sygno is at home at https://www.sygno.com/ but also on LinkedIn at https://www.linkedin.com/company/sygno/

You can learn more about myself, Brendan le Grange, on my LinkedIn page (feel free to connect), my action-adventure novels are on Amazon, some versions even for free, and my work with ConfirmU and our gamified psychometric scores is at https://confirmu.com/ and on episode 24 of this very show https://www.howtolendmoneytostrangers.show/episodes/episode-24

If you have any feedback, questions, or if you would like to participate in the show, please feel free to reach out to me via the contact page on this site.

Regards, Brendan

The full written transcript, with timestamps, is below:

Sjoerd Slot 0:00

Holland has always been a great point for startups to try out new things, while at the same time you're being forced to look internationally.

We just had to acknowledge we were not getting up to the speed of fraudsters evolving and money launderers evolving. The tendency, then, is of every human being 'so I see the fraud, I'm going to stop the fraud because I don't want that fraud to happen again tomorrow' - but they were getting smarter, faster. And so that was a big turning point that 'okay, we need to change that approach'.

And you need a number of frauds before you figured out what you're doing, but fraud is only 0.1% of your actual transactions so that means you throw away 99.9% of your data. And also, game theory says fraudsters are going to change their behaviour, they're going to try to circumvent your rules, they're gonna adopt their behaviour while 'normal' customers want.

So how do I switch it around? How would I found this type of fraud while not looking at the actual fraud itself?

Brendan Le Grange 0:50

In some ways, my whole career was shaped by a decision someone in HR made before it even started. I was part of a graduate recruitment intake, one of a dozen like minded kids chosen to help build out an experiment Capital One was running in South Africa. Now, we really could have been shuffled up and spread out in any order, but apparently, at the very last minute, my assignment was switched so that, instead of joining the scorecard building team in the mass market loans division, that went to my water polo playing, male modelling on the side colleague, and the credit card fraud team got me.

Cue some light hearted disappointment from some of my teammates, and for me a career that was close enough to scorecard builders that I could eavesdrop on their chat, but not so close that I could do their job for them.

As such, so some of you know better than me, but I'd always simplified the concept of modelling by saying that the target variable should be the thing that is uncommon. Most people will repay the loan, so we model defaulters; most people in collections don't pay, so we model propensity to pay; most transactions are legitimate, so we model for fraud.

But if you think about it for a minute, fraudsters are intentionally elusive. They actively change their patterns of behaviour when the old way stop working. real customers don't do that. So we're aiming for a moving target. When a static one is right there.

Sjoerd Slot, today's guest, thought about it and flipped the transaction monitoring model on its head, welcome to How to Lend Money to Strangers with Brendan le Grange.

Sjoerd Slot, co-founder of Sygno, welcome to the show.

Sjoerd Slot 2:43

Thanks, great being here.

Brendan Le Grange 2:44

Sygno provides automated machine learning software for transaction monitoring, so a fairly technical niche. What did your early career look like? And how did it set the groundwork for what you're doing today?

Sjoerd Slot 2:57

Well, I've never had any linear line in my life. Even going back to school, you know, I was great at math, horrible in languages and dropped out, sort of came back in again; same thing with university, I decided, you know, I want to change the world. So I ditched all my math and went to sociology and anthropology.

Figured out I wasn't really good at that, so finished it at some point in time, did a lot of stuff: traveled the world, was active in a student organisation called AIESEC, which built my international connections.

Then I decided to still want to change the world when the United Nations wrote economic reports about the role of youth, or the lack of the role of youth. Nobody was reading those reports, or they were reading but no one was acting on it - what am I going to do with my life?

I came back to Holland, didn't have a job didn't have anything. Well, you can always go into the financial sector. So, I think that's actually how I ended up in my career. I ended up in consulting and finance sector, and then quickly actually figured out that the whole thing behind the finance sector is all about trust, you know, that's also your podcast - that if we could trust the other party, we could just put money on the table.

So if you can figure out what money is going out, that shouldn't have left the institution, well, then you've sort of figured out the trust part there.

So with that in mind, I ended up in transaction monitoring, went from one project to the other, and then, at some point, figured out that we're doing the same thing over and over again. It's a great consulting model - you can charge high rates, there's always crisis, and you're doing the same trick, but you're not really helping customers there. Yeah. So I think that sort of also ended up with me starting Sygno. I started thinking about okay, so where did I end up now? And what are we doing here? And does this make sense, if you just take a bit of distance from what we're doing here?

Brendan Le Grange 4:33

Yeah, because the strategy consulting world is good at training you to think strategically, to think broadly, but it's still a big leap from from telling people how they should run their businesses to actually starting your own business - so what was that spark that got you to say, I'm actually going to go into it myself?

Sjoerd Slot 4:51

Yes. So the route, of course, is typically when you're a founder coming from a career part is that you have an idea that nobody else is doing it and then, at some point, people are telling you, 'well, maybe you sstart doing it'.

And that's the actual thing. So I started with a couple clients and we were just thinking about, how do we solve this problem.

And you know, this was the time when fraud was really hard to detect. In the past, you could have fraud rules live for a couple of years, you know, in the credit card space, in the insurance space. And the same with money laundering, you know, those rules would exist for five or 10 years easily.

But we were chasing, you know, man in the browser attacks, man in the middle attacks, and stuff like that, skimming attacks that were moving countries, so the speed of evolving was so fast that the old school approaches weren't working anymore. And so my client was desperate, saying, well, I'm losing a lot of money, or my clients are losing a lot of money and we're reimbursing them, and how do we protect them?

I said, well, every week we're adjusting the rules. And every week, we're trying to figure out what they're doing, we have big forensics teams figuring out what they're doing, the only thing we can now still do is try to be ahead of them.

In the end, they're gonna do something that doesn't make sense for a normal customer to do. So if you monitor on that, I think you can actually get ahead.

And so my customer says, well, do you know how to do it? And I said, well, I don't know, but I can ask around, see if there's some smart folks around, give me a server, give me some data. So I hired some Harvard guys. And we just started a project and said, 'let's try out on 12 terabytes of data, to see if we can actually figure out how to detect frauds. And we got it to work.

Of course, not in a productize way, you know, in a project way, but the idea works. And that's sort of what we thought, well, this might actually be something that we want to continue on.

Brendan Le Grange 6:20

And I don't want to jump too far ahead at once, but my first job was in credit card fraud detection and, at one stage, we were being hit by a syndicate - we had just launched the first major premium card in the market, we were very proud of it... and it started getting hit by fraud. It was the very early days of online fraud, so technically we weren't losing any money, because we could charge back it all, but it was happening so fast that these are premium customers of the bank, we would phone them and say we've picked up some fraud on your card, don't worry, we've closed the card, we're sending you the new one - and back in those days it would take about a week to print out a card and get it to the customer, and in that week, we'd have to phone them back and say actually, sorry, the new card also has some fraud, too, so we've closed that one and we're sending you another new one.

And the impact just from the brand was horrific.

And it wasn't very complex one, I won't go into sort of how we fixed it quite easily, but suddenly, we're moving from a world where fraud was somewhat random, it might hit one customer every 10 years to okay, we've been hit by somebody who understands the system, and the system is too inflexible to fix it.

We just didn't have the capacity to model data like we can today. So that's something I'm going to pick up later on on how you've come around to that. But first, I want to talk a bit about that founding experience, what was the FinTech world like the startup world like in the Netherlands 10 years ago, and what you see it like today, as a founder, having been part of that journey?

Sjoerd Slot 7:48

A lot of things are still the same. Holland is too small of a country to run a big business in, but large enough to run a good test environment in - so you can actually test things in a market, it's a mature market, great technology, people adopt technology. And so Holland has always been a great point for startups to try out new things. Because it's small enough, if you sort of hit the wall here, you know, you're not losing reputation on big markets, you can try things out, you can pivot, and they're going to adopt it.

While at the same time, you're being forced to look internationally, we do a lot of business in the US right now. If you're a US startup: Europe, you wouldn't even consider it, you know, it's too much of a hassle, etc. For a Dutch startup, you know you're gonna have to go abroad, you're gonna have to look around!

Same thing with international connections, conferences, you know, people love coming to Amsterdam, so therefore, we have a lot of conferences. So a lot of times we can go to comps, without having to have all the costs of travel, and the time and so that that's a great part there.

But yeah, the evolution we've seen is a lot of accelerators coming up, I think that's a good thing. Historically, a lot of those accelerators were connected - and I'm going really back to the beginning of the century, I was at the entrepreneurial university Twentw, we had these great staff and you know, guys from booking.com came from there and the JustEat takeaway guys are coming partially from that ,was a great time if you if you look around which great businesses evolved from that, but if you then look at my type of startup, which is a professional coming out of professional career, that type of accelerator wasn't existing, you know, we really just had to figure the thing out. I think there's much more accelerators now for more mature founders taking a different risk, but also take more experience in terms of knowing the market understanding the market, which of course, when you have a regulatory compliance, it does help you understand a bit what a regulated thinks like!

I think that has really matured, as I've seen a lot of folks I've talked to, in the last couple of, you know, two, three years, which have gone into great programmes like and learn you know how to really helps founders like that, to really get started when you're not, you know, a university student that is just going for it. Because I think we're done getting startups to work in mature markets where you do need to experience you do need to understand really deep domain knowledge that is present there, but still getting people who think slightly different, how have different ideas have different approaches and actually can solve problems.

So I think that's the biggest leap forward.

Brendan Le Grange 10:00

Yeah, yeah and, well, London's crown is always at risk, thanks to some of the voting public. But anyway, a lot of good news and good energy around the the startup FinTech world. But let's come back to the main subject: know good, catch bad.

Talk to me in practical terms about what that means, and I guess what inspired that philosophy that flips the traditional credit modelling that became the traditional fraud modelling approach on its head?

Sjoerd Slot 10:29

Yeah, like I said earlier, a couple of things really triggered it.

First of all, we just had to acknowledge we were not getting up to the speed of fraudsters evolving and money laundering is evolving, we're getting smarter, faster, we were closing down money leaving to certain countries, within two weeks, they had money mills in the Netherlands. And we suddenly had, we had some we had a domestic problem.

And so that was a big turning point, that okay, we need to change that approach. And you need a number of frauds before you figured out what you're doing. So that didn't work either. And the same part is we were struggling with legacy systems. And so a lot of that legacy stuff didn't have the data that we thought it was needed to detect, you know, those are the best for us.

So we said, well, you know, you can't just create data out of air. And so then we said, well, what happens typically is when you have a case, a case manager will tell you what's different than what you expect for this customer in this situation. The tendency then is of every human being to say well, I see the fraud, I'm going to stop the fraud, because I don't want that fraud to happen tomorrow again, and I still have customers asking me, How does your approach help me detect that specific fraud that I had last month or last year? And we typically have done?

Well, I don't know, I can only say, overall, we're going to find far more than you're finding right now we're going to stop far more, we're doing cybersecurity school to zero day attacks, we're going to be very early on detecting new types of fraud, we're going to have far less false positives, because when you extrapolate, you know what you see in a certain fraud, the things that you're seeing there are going to be extrapolated to a lot of good customers as well, specifically, I think the ones you're mentioning, you know, the more high end customers doing exotic types of behaviour for them.

That's normal.

But you know, other people travel internationally as well, other people do high payments, etc. So how do I switch it around? How would I found this type of fraud while not looking at the actual fraud itself, so it's really disciplined in terms of modelling.

But if you switch it around, old-school game theory says fraudsters are going to change their behaviour, they're going to try to circumvent your rules, they're gonna adopt their behaviour, while normal customers want, they have an intrinsic motivation to do that type of behaviour, because that's what they want to do. They want to transfer money to their grandmother or to the grandkids, or whatever the opposite side is also fault is only 0.1% of your actual transactions.

So that means 99.9% of your data is about non fraud. So if you want to detect fraud, you throw away 99.9% of your data. And we're looking quite often those projects about our data enrichment data quality, because we're focusing on 0.1%. But when we started, we said, well, if we focus on 99.9% of the data, we don't we still have data quality issues. And we would always like more data, but a little bit of data on 99.9% of date is far more than a lot of data on 0.1%. And so when you switch that around, it suddenly makes sense.

And in cybersecurity, we see this you know, network detection, fires addictions, are the enter any of them are doing fingerprinting on the actual viruses, or attacks. They're all doing fingerprinting on what is the expected behaviour for this device to be doing?

Brendan Le Grange 13:14

It just makes so much sense that if I've got two different fraudsters both acting on my account, you could be asking how similar is the activity of those two fraudsters potentially entirely unconnected, but both are different to me. So whether or not they're doing something similar to each other, whether they're doing two completely different frauds, it's highly unlikely that that transaction looks anything like mine.

And yeah, it's something that we had seen before, as I said, I don't want to bring it back away. So transaction fraud in the credit card world, but I think it was 90% of credit card transactions happened at a merchant that the good customer had used in the last few months, and very little of the fraud, almost none of the fraud happened at a merchant you'd used before.

But we couldn't do it analytically. We just didn't have the computing power to do it on that what, from a technological point of view or data point of view needed to happen to enable you to process in reasonable time, the 99% of data, the full base of good customer behaviour, and to use that as the seed.

Sjoerd Slot 14:15

Absolutely.

So I think the advancement in terms of calculation and compute power has really helped us and also, you need to be smart and what kind of questions you're asking the machine to calculate for you. And this is also what we do, we tried to do pattern recognition. So we first tried to figure out the pattern before turn it into a model, the faster we can do it, the easier we can do it on the larger volumes, we can do it and that really helps.

But the other hand is also if you want to have fraud, fingerprinting profiles, which a lot of parties are trying to do right now. You have to share that information. You know, there's a lot of technological challenges coming in place. You have to keep up with that. I think that's also one of the reasons why in the virus scanner world you're not maintaining the virus libraries anymore because they just huge.

There's no way you can you can keep that and compute against that in a quick way. So the model generation is a very compute intensive because we need to do pattern recognition, we need to figure out what is this data all about what's happening here, who are the customers who are the entities acting here. But once we've calculated the model that you actually put into your projection onto our system, we typically reduce the computing power needs, because it's spot up.

It's not something that people thought, well, this might work and then we make an exception rule, and we make yet another research, we build rule upon rule upon rule, which becomes very slow.

Now we, we use a lot of computing power in generating the model. But once the model is generated, that's the spot on model that actually works. Yeah. And so that has far less steps in it than the average, you know, organically grown fraud detection rules. Base.

Brendan Le Grange 15:37

And let's talk about what it looks like in a product because I don't want to derail this with my reminiscing about my fraud days, what does the signature product look like it? How's it used by your customers? What industries is it used in?

Sjoerd Slot 15:49

Yeah, so it's using fraud and anti money laundering anywhere where transaction happened. So we have customers in the historic world of card issuing and acquiring wire transaction online banking, correspondent banking were currently talking to and working with customers in funds and trust businesses in the payout of insurance claims. So that's we're still talking about the actual payment transaction, not so much declaims detection, and has been used by banks, payment processors, core banking processors, everybody was already sitting on that payment, say, well, I want to add to service for my customer or for myself.

Brendan Le Grange 16:23

Now, they can sort of be a question of, well, can we see what's happening? Do we understand what's happening? But you very big on transparency, I see other website, but in the fraud world, then we say, well, if the model is transparent, are we showing fraudsters our hands? Are we giving too much away?

So how do you balance the need to be transparent to have rules that the regulator can see and understand and be happy with, with having systems that are protected against fraudsters?

Sjoerd Slot 16:49

Well, let's go back to your example you were giving about, you know, we don't see a lot of fraud with merchants that the customer has already been with, we do see that, because what happens is when you when you harvest the card, you're gonna make a very small transaction that the customer couldn't be bothered about, I think what is this, but I'm not going to make a fraud report about it was just a couple cents. Yeah.

And suddenly, you see a week later, or months later, you suddenly see a big harvesting of the cash.

So a simple rule would have said, if the customer has already been at his merchant, you know, let's not look at it anymore. That's currently that's one of the main attacks that we're seeing. So what is important, we always talk about models, those are combination of rules that need to work in conjunction with each other.

So it's not one rule, if I just circumvent one rule, I need to be able to serve multiple rules, and indeed stay on the multiple radars, which is already more difficult. And then we just those radars, to the individual, what we call entity behaviour, that can be a customer that can be an account, that can be a car, it can be a terminal, that can be a merchant, I'll give a very simplified example, if you stay on the the average of this customer, but your average is going to be different than my average is going to be different to everybody's average.

So I need to figure out on the individual customer level, what is normal for this customer, probably against an individual merchants on the other side of the transaction. And so then it becomes really difficult. And game theory tells us if it's difficult to predict, I'm just gonna go for what's easy to predict, I'm gonna go somewhere else,

Brendan Le Grange 18:05

When you are applying these new techniques. Is it possible to have Explainable AI in this space? Is it possible to have rules that are on the cutting edge in terms of the technology, the modelling that we have available to us today, but then we can actually explain to a regulator or to a customer, what it's doing,

Sjoerd Slot 18:22

You need to look into what are you doing with the AI? And what are you coming out of, and I think one of the great benefits for us as we come out of the market, but we've also started with, you know, we want to work with legacy systems that our banks and payment processor working with, and because if you want to change your transaction monitoring system, you know, you're going to take two or three years to do the project. So we said, we have to work with a system, those are typically rule based or very lightweight model type of system.

And so that really helped us also in the explainable part because when we say explainable, we also mean comprehensible, you can, you know, there's a lot of, you know, shapely value type of Explainable AI things you can do, the difficulty is not so much explaining it in terms of that and other data, scientists can understand it, the difficulties that a legal compliance officer, it has no mathematical background, you still understand what you have. So then we say it has to be comprehensive.

Also, if you cannot explain to your grandmother, so we use all the AI and machine learning to do the pattern recognition to make that lightweight model, which is an Indiana is just a simple rule set with a scoring and decisioning rule on top of it. So we use a very computing intensive machine learning intensive approach to come up with something that is pretty straightforward. And you can actually print on a paper and we have a rule in our, in our development teams. If you can print a model in less than four pages, it's not comprehensible anymore, people are gonna lose track of it. And so you have to come down to Okay, so how do I use machine learning to do all the pattern recognition in the data?

But I need to normalise it at some point to something that I can explain to somebody who doesn't have a mathematical background and says, Well, I wouldn't have come up with it myself. But it does make sense just like the old fraud expert would have come up with rules that the average business person thinks, Well, I wouldn't have come up with it. But I can see your ex parents hear you come up with some intuition thing.

Brendan Le Grange 20:02

Another thing I saw in that AI realm on your website is a focus on these unintended biases. Now, it's a bias in AI or in modelling is quite closely linked to explainability. So what what is the philosophy at Sygno around working around bias? And how do you make sure it's not accidentally in bolt.

Sjoerd Slot 20:24

So when you look at a lot of the financial crime approaches, again, coming back to you have a very limited data set, you know, limited crime, that you're extrapolating on limited parameters to the whole audience. And that's when a lot of these biases come out.

And the Dutch regulator just published a new statement last week, together with the Dutch banks say we want to stop the de risking of the bank. So they are they've been cutting out, you know, old types of associations, types of businesses, that there were high risk, because there was a couple of frauds or a couple money launderers in that business. Yeah, of course, there is more risk in certain businesses than others, we know that a lot of banks will say, Well, I don't want to do business, because I don't want to get to find it, I want to get the remediation programme.

And that really leans on the approach of saying, I see frauds identify fraud or money laundering. And I'm going to extrapolate what I'm seeing to all the other businesses in that same domain. And so a lot of the biases issues, there's always bias in it, because that's what machine learning does. But the biases that are getting issues are those type of issues are saying there's a certain type of basis we've seen throw out there.

And we are not able to distinguish between the fraudsters and a non fraudsters in our business efficiently or effectively. So therefore, we're just going to rule out all these are people making transactions to certain countries, because that's our home, country, etc. And so when you turn it around, so what if I can go on an individual level saying, this makes sense for this individual, this is probably legitimate behaviour of this individual, you've made transaction, this country or you're in this type of business, or you're this type of person or living in his neighbourhood, etc, I can certainly go on the actual individual behaviour and say, well, this does make sense.

Brendan Le Grange 21:53

And especially when you think about the levels of risk that make banks nervous, the banks are going to take a 10% risk on on a loss. So it might still be nine out of 10 people in that neighbourhood or in that industry or that country, are perfectly good. Being able to see past the sort of redlining is is a massive impact on so many people's lives. And specifically

Sjoerd Slot 22:12

in the ethical side is where you want to enable it, people that want to do good, but just come from a bad background, or have a bad name or have some, you know, something that at some point, somebody thought, Well, I see a fraud, then I'm gonna look at the very obvious stuff out there, or maybe, you know, they thought it was obvious, and I'm gonna just mark, everybody has the same characteristics.

Quite often, it's 99 out of 100, that are doing good, and it's only 1% as bet, then you're already quite high.

And so you want to make sure that all the other people are being able to do the business that they want to do or make the transaction that they want to do. And it goes both ways. It goes for people who are underprivileged, as well. So people are privileged, I've been travelling also to New York or Washington with some folks who used to work at a bank. And the first thing they did when they arrived to JFK was go to the ATM and take out cash. And I was like, why you guys taking so much cash out and go, like, I have three credit cards already on me but I know when I'm going, I haven't been to the US for because we're talking post COVID, you know, I haven't been here for a long time. And I know the mechanisms are out of my bank operates because I used to be responsible. I know they think you haven't been to the US for the last two years due to COVID. So we're gonna block that card. And even if I have three cards, probably by the end of tomorrow, all my three cards are going to be blocked. But this is totally normal behaviour for me because I am an international traveller, I have funds etc.

So the strange thing is, a lot of false positives are hitting the underprivileged and really privileged ones. And for a bank, one of them is unethical, but also a business opportunity side because there's there's great people that want to do the business one, do the transactions. And the other side is also a lot of business for a bank or a payment processor. That's the people that you want to have in your in your portfolio.

Brendan Le Grange 23:49

I guess 80% of people are in the middle, and the either 10% or getting caught. But the 10% on the top end, have the ability to phone and yell, yell at someone in the bank and get it fixed but the other 10%

Sjoerd, your website is one of the nicest design ones. I've seen really cool animations and also just lots of good content on there. Where should people go if they want to try and get their heads around? The I know, there's some videos on there that explain the concept as well.

Sjoerd Slot 24:16

The website is www.sygno.com Always great to be there.

I think we've tried indeed to look slightly different because we have a different approach. We also don't want to go to the market with you know, old school, blue orange type of proposition, and also to trigger people thinking, you know, what are we doing this market? And what can we do differently?

So go there, there's great content, follow our LinkedIn page, because there's some great content coming on there be part of that discussion, we are at events as well or people can approach me or my colleagues directly.

We are at the station in where the market is making a shift. And this is an interesting time. It's great for us to be part of that shift. And I think we're really seeing it where you know, US regulators pushing for fraud today email integration for Model Management for how do you take the unknown risk the ethical All parts have all those false positives that are being generated.

So it's a great shift where the market needs to move away from the old school: I've seen a fraud so now I think I know all fraudsters to, I've seen a good customer now I know what a good customer looks like.

That's a great conversation that we want to be part of. And we want to hopefully lead but definitely be an active member.

Brendan Le Grange 25:18

Yeah. And I want to close picking up on that point, from your experience working right there being one of these changemakers. What are the trends that you're looking at the moment that those that are not as directly involved should be keeping an eye on?

Sjoerd Slot 25:32

First of all, the integration of fraud and money laundering, you know, in the end is money going from A to B that shouldn't have gone from A to B, whether it's fraud or money laundering, it doesn't really matter for a regulator? And you know, who's a monumental? Are you money laundering or fraud? You know, so that's a big part.

And then the whole, how do you detect the unknown risk? I think the regulators are really caught up with like, Okay, so the old approach doesn't work anymore. When the risk is evolving that fast.

So how do you detect something that you don't know yet? You know, what, what we call cybersecurity to zero day attacks?

And so how do you make sure that you have all these controls in place, but they're sufficient for something that you haven't seen yet. And the last one, I think that's the most recent one is Model Management. So in the past, credit card models could exist for five or 10, maybe sometimes even 15 years, or money laundering rules as well. But you looked at them, it's all going well, it's fine. Right now, the regulator is pushing for a model evaluation every 18 months, but probably coming down to 12 months, or maybe even more, if you look at the average model projects here.

So we of course, are software that generates those models so that that's a couple hours of compute time. But in most of the banking world is still project based, you know, human based data science, those projects last longer than 12 months, and those backlogs are two to three years easily. If you didn't say, well, you're going to make a model, but you're going to reevaluate it every couple months, those number of data scientists cannot be recruited in the market anymore. So that's going to require a whole different type of approach,

Brendan Le Grange 26:55

it also speaks back to that idea of of modelling the good because we put a model in place, it's turning away bad. So now we're filtering through, hopefully, far fewer. If you need to rebuild it in six months, you don't have the, you know, the data stock of bad customers to model. Whereas if you're doing the good customer, you've got that multiple of 90 odd times more data,

Sjoerd Slot 27:17

and the good customer is going to be more stable in their behaviour. So where the fraudster probably has changed in those six months, the good customer does a little bit, you know, we've worked through COVID, we've worked through the current inflation, so there is definitely changes in behaviour. But overall, our models are very resilient against those those changes, and they handle zero day attacks far more efficiently than then the old school fraud rules.

Brendan Le Grange 27:38

I think also just in closing, from a consumer point of view, it makes for a far more understandable impact. If you do get caught up in a process, if that is related back to me and say, Well, you don't normally do this, versus the other approach where you fly to America and you Your card stops working and say, because you're in America, and you say, Well, yeah, I've been here 10 times, even a country you've not been to before, you should be able to fly to Canada the first time and say, Well, yeah, I've not been to Canada, but I've travelled to four, five countries. That's where it gets frustrating where you say that you've got everything on hand to see who I am. But you're blocking me that even if you were to be impacted as often, you would rather be impacted for behaviour That's unusual for you then behaviour That's unusual for the average.

So yeah, it's really interesting space.

Also just exciting for me to see where it is compared to what it was 20 years ago that there really is just as much data innovation data work happening in the back end as there is or maybe even more now than then in the front end. It should thank you so much for coming in.

Sjoerd Slot 28:39

Great to be here.

Brendan Le Grange 28:40

It's been really interesting to hear what you're doing at Cigna.

Sjoerd Slot 28:43

Thank you very much. It was great to be here.

Brendan Le Grange 28:45

And thank you all for listening.

Please do look for and follow the show on your favourite podcast platform and share the updates widely on LinkedIn where lending nerds are found in our largest concentration. Plus, send me a connection request while you're there.

This show is written and recorded by myself Brendan le Grange in Brighton, England and edited by Fina Charleson of FC Productions.

Show music is by Iam_wake, and you can find show notes and written transcripts at www.HowtoLendMoneytoStrangers.show and I'll see you again next Thursday.