Oscar Koster and big data scoring for thin-file consumers
The traditional credit model is often underpinned by an existing credit history. This makes sense mathematically, after all the best predictor of future delinquency is past delinquency, but it can present a barrier to entry to some customers – if you won’t give me credit today because I haven’t had credit before, then how can I ever get credit?
Consumers with thin files - or indeed no files at all - on the credit bureaus found themselves all lumped together and burden with a high-interest rate. In the big developed markets, this might be a small population and resolved by one or two lenders taking a risk. However, in developing markets many lenders don’t fall within the bureaus’ catchment areas and so even borrowers with a good history with credit may not have a bureau file that reflects that.
This is much harder to resolve. Or at least it used to be.
In today’s episode, I speak to Oscar Koster of BigDataScoring.com about the ways in which they are using alternative data to create predictive credit scores in developing markets. From social media connections to satellite photos of nighttime light production, the modern world is a rich source of data if you just know how to use it.
You can reach out to Oscar by email
This is also the first cross-over with my other, temporarily paused, podcast – so if you’d like to listen to Oscar’s thoughts on mountain biking for mental health, head on over to https://www.themostfunyoucanhaveonabike.com/episodes-1/oscarkosterandthepostridebeer
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 time stamps, is below:
Oscar Koster 0:00
There's a whole bunch of people out there where the traditional model doesn't work, there simply isn't enough information on these people to make a reasonable credit call...
Brendan Le Grange 0:08
Welcome back to How to Lend Money to Strangers. In today's episode, I speak to Oscar Koster from BigDataScoring.com Oscar Koster is a Dutch mining engineer, turned South African banker, turned international entrepreneur. I met him at the midpoint of that progression when he was my boss, most famous for organising a team bonding day, where he took us to his previous employer - the explosive factory. But I haven't made a podcast about how to efficiently break and move a rockface using a series of carefully placed and perfectly timed charges. How to Lend Money to Strangers is a podcast about lending strategies around the world, and across the credit lifecycle. Sometimes those strategies are underpinned by traditional data and tools and sometimes, like today, there's something a bit newer on the market.
Your entrance into banking was an unusual path, but in terms of your experience within lending it was working with credit cards and working with debt collection, and traditional scorecard models and a market with multiple credit bureaus, solid data, and where you can work a traditional scorecard approach. But now, what's the approach you're taking?
Oscar Koster 1:39
I'm focusing on an Africa, but BigDataScoring the company, is largely focused on South America. It'ss really big for them, the management is based in Chile. But in general, the developing world is seen as the market we're going after - it's traditionally the people have been excluded from the credit cycle. There's so much data available on, say, Western Europe typically, that you don't really need to resort to alternative data to get a image of what someone is like. There's there's a lot of stuff available.
Brendan Le Grange 2:13
We found, within the credit bureau world, that often the fact that we could tell somebody was definitely thin file, definitely new to credit, was the most important thing. So you could build a new to credit score, but actually the most important part of your new credit score was the fact that you didn't have bureau data. The vast majority of people are there and so the fact that somebody isn't, tells you something, and you can infer quite a lot from that.
Oscar Koster 2:39
So what my entry into this was really, after I sold the debt collections firm, and there, we pretty much tried to work that along, at the time, to apply data in a way that you can work smarter on things. And that sort of way of thinking of what else can you do with data, that sort of stuck in my head.
And that's when I started looking around here, well, firstly, I guess close to home to South African context, clearly, we've got a large thin file population - not necessarily unbanked, but thin file. And when I say not unbanked, that might be things like Thyme or other little things, but this, well it tells you something but it doesn't tell you everything either. Combine that with the pressure of financial inclusion, and that sort of made me realise 'well, there is actually something to be done there'.
I then stumbled across, initially, Jumo.world. At that stage I didn't know what, but they largely use cellular data to build credit models, and they're lending up in East Africa specifically. That made me go, well that's pretty interesting, but who else is playing in this market? And that's when I came in touch with BigDataScoring. So effectively when you ask what 'what are the markets we're focusing on? Yes, it's a relevant market, because we feel we can make the biggest difference there. Therefore, that's also the best market for us to go after. Yeah, I'm sure we can improve scoring in a more established models, but it will be marginal. I mean, if you do a good job modelling, you should be able to construct a pretty good model based on bureau data in the developed world.
Brendan Le Grange 4:07
Yeah, and I think in most developed worlds, there's already more segments that are possible to make, than they could practically put a strategy against. So, yeah, we often would run against that - we'd say I can give you 50 types of customers... I only want high, medium and low risk. I don't want to deal with 50 scales of risk. Yes, you'll find somebody who might have some marketing opportunities for that to get down a little bit finer, but from a risk point of view most of the time, the risk data exists.
And so when we talk big data, what sort of data are you talking about, in particular, like what are the sources of that data you're bringing in?
Oscar Koster 4:42
So we we do this in a couple of phases. The easiest phase is where we simply get an email address, a cell phone number, and a physical address. And based on that, we can then ping a whole bunch of databases - typically pick up somewhere between 3,000 and 5,000 data points per individual.
What's the general use of the land? Is it industrial, is it residential, is it mixed? How many schools are in the area? How many shopping centres in the area? How many parks and how far is this specific spot from any of those points of interest? Sort of intuitively, it makes sense that a residence which is fairly close to a park, a shopping centre, and a house is probably more valuable. So that's the sort of thing. Another one, we do is spectively a night typography of the amount of artificial light in an area: the Northern Hemisphere is lit up like a Christmas tree. For the Southern Hemisphere, it's pretty dark. And if you can start zooming this in on areas where people live, again, it's the link to affluence. That's the sort of stuff we do on address.
On cell phone number it's things like who's the provider. And interestingly, different providers do attract slightly different customers. So in a South African context, Cell C was a late-comer to the party and always had a value-for-money type of offering, but they had to lure largely existing cell phone clients over to them. As a result, the characteristics of the Cell C population are a little bit different from say, Vodacome, MTN. You can also see whether a number's been ported. Clearly, whether it's a prepaid or contract. And whether it's a company phone or a personal phone. Those are the sort of things on that. Email address is, again, is this a company email/ is it a personal email. Who's the provider. The other thing, which is interesting, and I didn't know until I started working with big data, is the actual makeup of the email address also has a link to risk: so if I take myself as an example, my personal email addresses is my initial and my complete surname. Typically, that is better in terms of risk versus something like oc 3268.
So what I just spoke it through is what we call phase one. And that's where we always start, because this is easy to do a PoC on. Give me the data, I'll construct a model, will prove to you by supplying you with ginis, or KSes, or however you want to measure, that's how good it is. And then we can do something with it without requiring massive integration, data flow up and down, all that sort of stuff. And in terms of our modelling techniques, it's it's, it's a combination of the classic ones, like regression analysis all that sort of stuff, combined with some AI. Traditional data tends to be a pretty limited set of variables, which have a very strong link to risk, that slopes it very nicely. Whereas alternative data is the polar opposite, in that it's, it's a plethora of different variables, none of each separate one has got a very strong meaning of risk, but they need to be pre processed to come up with a bunch which then have a strong link - but we start with 5,000 variables initially.
Then the next level up is, if all your client acquisition is digital, then instantly, you've got incremental information, namely, what's the what's the behaviour on the website? Do people read the T's and C's? How fast do the type? Device information - what are they using to access you in terms of what is the operating system on a device? What device is it? How many mistakes do they make? It's all that sort of stuff you can measure. If people use an app, then you can also start scraping stuff off a smartphone. Basically, you send someone a cookie to a smartphone and say, listen, you've applied to us, we would like to get access to the following information on your smartphone and people can tick things - whether they allow you that access, yay or nay. So they don't look at, you know, who does Brendan know, but I do look at what's your ratio between inbound and outbound calls. As you can imagine, that's really rich information.
It's also got some downside. Practically, it requires quite a lot of integration between you and your clients, a lot of PT and cost to go through for the sake of doing the PoC. Unless you come up with some generic scorecard, but, well you're a data guy so you know, a generic scorecard across all products, basically globally, is unlikely to be very accurate. I mean, it's just not specific enough to any specific population. Yeah, I mean, we, the typical improvement we see in gini is of the order of 25%.
The other thing we found, which is again that will speak to you also from your CapOne experience, is that by asking people to get access to all this information on your smartphone, you get negative self-selection. In that, who are the people who give you all the access? The super risky ones? Who are the guys who are saying, I'm not sure if I want these guys to know that stuff,? Those are probably the people with a slightly better risk profile. So if that is the only thing you do, you got to be careful not to rob yourself of your best potential prospects.
Brendan Le Grange 9:54
You know, in a world where we all told about being careful of scams, it does make you nervous. It's a similar problem open banking have had. But open banking, I think, can benefit by the fact that it can come from a big name bank that you've, you've heard for 200 years, you can trust them. But if I saw it, yeah, if I saw, particularly if it's a separate app, right, and you're like, well, it's credit..
Oscar Koster 10:14
At least most of us have got a white label, if you would apply with ABSA, you will get an ABSA link. But being that makes this integration tricky. Yeah, this is not some Mickey Mouse front you can stick on the top of it. And in our experience, although it's powerful, it's hard enough with... this is sort of not necessarily risk inherent but commercial, it's hard enough to speak to them to get them to do something new, without having to tackle all sorts of integration. You know what IT prioritisatin is like, there's always a bunch of stuff which got proven business cases of X amount of billion of Rand. And then there's this oke wants to do something funky and new - it's hard to get through the door even with it. Where they make the biggest inroads are digital lenders, where this stuff is relatively speaking, much easier, they've got the whole infrastructure set up.
Brendan Le Grange 11:03
Yeah, and I think a lot of those, the digital lender,s understand the inherent value of these data fields, but there's nowhere to look to, because all the history of predictive modelling in financial services hasn't needed that, hasn't looked at that. And you can do ground-up work and a data analysts can look at all the data and, in theory, throw it in a model. But there isn't that history of knowledge of how to treat these fields that are obviously valuable, but sort of traditionally didn't exist. I think as well, if you're going to tell a bank, okay, now, in your internet banking I'm also going to want you to add in these steps, it's gonna be hard for them, whereas a digital bank has probably got somebody who can find a way to make it smooth.
Oscar Koster 11:45
Which is also why you see, more and more telcos moving into the space, because they already got all this data. And they've definitely got the base needed to do appropriate modelling, they, I forgot, there's a name of a specific file that you can pull off handsets that's got a wealth of information. And I guess, in East Africa, that's why mobile lending is so big. There's a couple of reasons. It's also very easy to disperse funds.
There's a whole bunch of people out there where the traditional model doesn't work, there simply isn't enough information on these people to make a reasonable credit call. In those situations, there's two approaches. One is, well, let's treat them as super risky and people pay a super high price for it and then hopefully, now, the income model is strong enough to offset whatever bad risk we take on those. And the overarching opportunity is to sort of put that grey bits into more black and white, because you're now treating a whole bunch of people who are maybe not as good as the traditional people, but clearly not as bad as you're treating them right now. In general, bigger banks take the approach: if we don't know, we want to say no. And the use of alternative data is a way to try and - you'll always have a bunch of people where you don't know one of them - but at least, as I said, sort of declutter some of the grey in more black and white.
Brendan Le Grange 13:13
This approach solves two issues. So one, you get rid of all the data capturing problem. And so when we were doing lending in Africa, it wasn't so much that, you know, the data was impossible, the traditional data it just hadn't been kept, it hadn't been stored, you would have had to build out a whole lot of infrastructure and culture to get all that data into a central point to model anyway. And the phone's doing that and solving that problem. But two, kind of more on the soft side, normally as you say when when people are pricing alone, they go so well, that the price is really high, because it's risky. Where it's fully covered by a credit bureaus and such, your loan is more expensive because you've got some negative behaviours in your past, you've missed some payments, have gone over limit - so you have displayed some risk behaviours that mean we can't take that risk on you without charging for it. But in the thin-file/ new-to-credit/ unbanked space, the risk is something I don't know. So I'm going to charge you a lot of money, because I don't know a better way to do it. And this to see the innovation come in this space is really pleasing to see on that front. Because you're saying, well, we can know about this. There's all this information that traditional banks are just not using. And here's a way to do that and therefore opens up pricing and/ or gives the acquisition team somewhere else something to show their boss because again, nobody's going to take the risk in their career to say, oh, yeah, let's open the door to this population, because I've got a good gut feeling. Yeah, but if they got the data, with ginis scores and KS scores, you can just fully plot that against your existing credit score and say, well, here's the new to credit population, or here's our test population.
Oscar Koster 14:56
It's interesting how all bureaus claim to be doing some stuff there, but at the same time, because in some way, they should be ideally positioned. I mean, they, they do have more data to, to work with than the likes of fintechs like ourselves. And at the same time, some of them are pretty bureaucratic, I think Experian and TU are good examples of, it's a big corporate, they serve big corporate clients. And in fact, it's for the vast majority of the big corporate clients of those bureaus, this is not a space that's very big for them right now. So they're probably better off leaving it to the smaller fintechs to do a bunch of stuff. And that's, that's what you see.
So what we do find is that, I think, this is a good sort of space for cooperation between a data house like a bureau and ourselves in that, for us, the big attraction is that decreasing infrastructure, so bumping the score through to a big bank is much easier, but run through to you or experience infrastructure, then us having to set that up, because then we again, ran up against compliance. And although we're doing something with JD and there, it's that simple, but JD is happy to work through an API into their system. That's not going to fly with an ABSA or Standard Bank.
Brendan Le Grange 16:16
Yeah, I think there's two points on that. And one, one on kind of putting that score in the bureau, or a centralised thing, I think makes sense. We did some work with somebody in the space a few years ago. And, yeah, the problems were a lot around the logistics. And the fact is, you kind of wanted to first check the bureau and then only if they weren't on the bureau, you wanted to run the score, but it created an awkward process. Whereas if it was all housed in one space, that's much better. Two, the data fields, you talked about at least the identifying data fields are things that a bureau often has, so you can run it, and then you can create that easy switch to whichever score you want being produced.
Oscar Koster 16:55
And then they can go out and offer that to their clients. And we do some sort of revenue share model. That's the model that we working on here. That's also one of the nice things about being a small FinTech, you've got way more flexibility than being a big corporate. So if people prefer to deal with it through an already trusted source, like like, a bureau, people can do it, that's fine. But for specific digital lenders, that might actually be better not to do to a bureau, because ultimately, a bespoke model will always be stronger than some genetic model. Yeah, I mean, both models can work. And ultimately, what people do without score is, again, very similar to how people use credit bureau scores. And yes, some some will, base that decisioning process on that score only. whereas others will have that score feeding into some process, which has got some knockout rules, or maybe some further processing that they do internally.
Brendan Le Grange 17:55
Yeah, and so we've been talking a lot about proof of concepts and things so far, but in the in the business itself, globally, I think you've 100 million credit decisions made already last year. So this is not all theory is actually...
Oscar Koster 18:11
This is a space where lots of people are working, but very few people can claim results. Because this is also the sort of space where lots of AI propellerheads think they can crack the problem. To some extent, that's true. This is also the classic case where progress is both hindered and aided with experience. It's actually good that some youngster on a beanbag, with long hair, thinks about this stuff completely unhindered by any previous industry knowledge, because that's anyone with too much experience probably thinks too much inside the box. At the same time, with something like credit, you do need to have some other people in there who can say, 'well, yeah, that's cool but you need to take these following five things in'.
That doesn't mean that the thinking needs to be restrained, but someone needs to make it practical in the end. To simply let the same space cadets go mad on this is likely to land you in a heap of problems, if you don't actually understand the lending industry.
Brendan Le Grange 19:12
Yeah, so we did some work with a FinTech to using alternative data to pull the score in one of the markets. And what was interesting that they took an approach, which on the one hand is really admirable. They went in first with a big pile of money, and they lent it out to a whole lot of people. And then they modelled it from there. And then they brought us their score. And they approached us and they wanted to do some work, and we took the bureau score. And what was interesting is that their score is really predictive on their population, but it became less clearly predictive on the population as a whole. And they - this is all then my speculation - but I wondered, in going out and lending as kind of unheard of FinTech people take the money from you, and then probably don't care that much about paying you back, especially since this was a limited time project so at some point, they just disappeared. So I wondered if, because they lent, and they got a whole lot of bad's that just genuinely weren't really that bad customers - but because of the nature of the business, they went bad.
Oscar Koster 20:15
That's assuming that people who took out that loan, we're aware of that nature of the business,
Brendan Le Grange 20:21
To some extent, but I mean, they were certainly aware that this is a company they've never heard of before, who is now giving them a loan. Their model, the score, and it worked really well on the data, but it was really over fitted to a population that didn't translate very well. Now, as a proof of concept of the technology, it was fine. But what it didn't do was build a score that you could implement, until you said, I get you got to go back, I now trust you can do this. But I think it fits to that thing where you can get caught up in the numbers, you can run away with it and then find, you can create these, these business models that are fantastically accurate but don't stand up to the rigours of the developing world.
Oscar Koster 21:00
I agree. So that's why it's when someone's literally not done anything in thin file, then, then you don't basically have the target variable to model. And that's an issue, the vast majority of people we deal with have have done something and would like to expand on this in a more accurate way. And that's normally what we deal with actually. No performance data. I guess the other one that you pointed out earlier, sort of the classic, low and grow writers indeed. Throw a bit of money out there, see what happens. And then, nice and quickly, update your models as you see performance coming through on these people. Yeah, it's an interesting space, because there's always new data becoming available, and it's gonna become more and more important, and banks are starting to look at it.
And I guess telcos is going to be potentially the interesting one, because they will probably claim that they can do quite a lot of stuff themselves, and that they have quality information. Some of them might also have analytical capabilities to do this. You clearly got the big prospect pools to go after as well.
Brendan Le Grange 22:12
Yeah. So that's interesting, because I was just talking to an ex colleague in China the other day, and if you look at how WeChat and Alibaba grew, they initially started, because they had all their data from their their businesses, they started doing lending to thin file, no hir type populations that the banks traditionally wanted to avoid. So they said, we can actually identify risk in these populations that you can't see. So we'll give them loans. And then that growth was really strong. And so then the traditional banks said, well, actually, you can help me convert some unproductive capital, I will give you the capital, and you find the customer and lend to them. And that powered the growth that we see. But now those banks have essentially taken all the risk of the customers they want to avoid in the past, because they funding the loan. But all the upside or the brand building is being done by Alibaba and WeChat.
And so now it's five, nearly 10 years into their existence. Those young no hit customers that they started with, when they were 20, are now becoming close to 30. They're going to look to buy a house. And when they go and think where am I going to get the loan? It's not Bank of China anymore, because I never heard of it. Even though Bank of China has been giving the money or whichever bank, whichever bank has been, or the upside has come to to reach out to Alibaba. And now, you know, there's some regulatory pressure there now, which will mean this won't happen exactly like this. But the banks have given away their business. And I think if the banks look at Yeah, the telcos could do it. And we could have the telcos help us. Yeah, they've got the prospect pool, they've got a few million customers, we can use some of their data, in particular market with is developing and developed alongside each other. And you're saying, well, these are thin file, you do run that risk that you might partner with somebody for too long, that by the time you say, okay, now this customer is mine, I want to give them a mortgage. They've never heard of you. And there's a mortgage FinTech that started up and giving the mortgage away.
Oscar Koster 24:04
And what you often see as the next step, then is that the banks end up buying that FinTech.
Brendan Le Grange 24:09
Yeah, yeah, yeah.
Oscar Koster 24:11
I mean, you've, you put the capital behind making something great. And then you end up buying out what you assisted making great. And maybe that's not such a bad thing, but that that's the typically reasonable deal.
Brendan Le Grange 24:26
It's typically a reasonable deal if you're getting in early enough in the process. So now when you look Ant Financial, you know, is essentially the fourth or fifth biggest bank in China. And I think regulations are going to step in so this decoupling of risk, where Ant was creating the loan but taking none of the risk, the regulators have cracked down on so it's not going to be able to keep growing like it has been, but the banks can't buy them any more. You know, in a normal market, you buy the FinTech a bit sooner and things...
But I think in a market, and I'm thinking more now East Africa where that big push has been telco driven, you know, Safaricom where behind the revolution in lending there. The banks are in a position where, you know, the banks in, in, in Kenya at least traditional banks, they're not gonna have the fleet feet that a FinTechs gonna have. And they will have to weigh-up that thing. If they don't do this, if they don't say how do I use alternative data and lend to new customer pools in 10 years time, the only people borrowing from the big banks will be the same ones that are borrowing now, so it's gonna just be a decreasing population. And, you know, I got the bank account from my dad's bank when I was a kid. And then you had the teenager card, and then I got my student loan from them. And then I got my first job, and I haven't been accounted their bank used to have these long relationships with banks. But I think there's a risk that you can partner away that if they're not getting into the space, if they're not saying, I need to lend like a telco would, I need to learn like a FinTech would.
Oscar Koster 25:55
In lots of ways., I think, dthe Far East aas been head of this the most, TWhere the risk you just mentioned , ofr the big banks, you know, being at risk or effectively selling, providing Ccpital Ofor these small operators to grow, and ultimately finding themselves having basically funded the sale of their own customers over time, plus all these fintechs Yeah, I might start off in in the sort of new to Mmarket or thin file , or whatever it is. But as these people got repeat loans coming in, I guess llikes fo Wonga, and other of these payday loans are a good example of that. They I don't think that will ever compete in terms of price of loan by the banks.the initial gap definitely gets reduced somewhat as people move their way up to, you know, from proper thing file to a little bit less than power.as they go through the cycle.
Brendan Le Grange 26:49
So I did a study when I was in Hong Kong. And it does involve a few assumptions but essentially, what I looked at is, here's the loan somebody opened up today, is it with somebody they worked with before someone new. Kinda has there been attrition. And then I looked at the price of today's loan, versus the market price for their risk, which involves some assumptions but basically the fair price, or is it based on the historic price? And I looked at the curves, and it suggested that people were more moved by how far the price for loans today compared to the last loan, then the market price, so they won't they're not aware of, necessarily, oh, I've got a super prime score, I should get 5%. They know that they paid 10% last time. And that makes sense. I mean, if you pay 10%, last time someone's offering you 8%, you'll take it even though maybe the market price is 5%. And so yeah, with all the price comparison websites and things that makes it a bit easier for people to be aware of the genuine best deal. But you're right, once people are in an ecosystem, as long as it's reasonable. And it's easy for them. They're getting cheaper prices, they're happy with that. They've got an easy process, because just clicking once you're not going in, you're getting the loan from them. You don't want to, you're not sure you'll get it from another product. But yeah, I think the broader principle is that it's not so easy to sit back and wait.
Oscar Koster 28:21
Some of this stuff naturally sits in the sort of information bureau world where it makes more sense to me ultimately, that's where a lot of this information sits already. It's interesting how you get different views. If you speak to customers of bureaus, you're being told that 'ah, the bureaus are also working on it' then you speak to the guys inside the bureau and it's like, 'yeah, we're working on it but....'
And when I spoke to the Kenya guys, they specifically listed TU and Credit Info as the bureau's operating. But it seems to me that Credit Info has got a lot more of the alternative data slant than the big ones. But at the same time, the fact that they're keen to talk to me about a bunch of stuff also tells me that their either not that far with it yet or they sort of have the sense that, you know, maybe it's actually better to let these dedicated FinTech guys, let them strut their magic stuff. And then after a while, as a bureau, I think you could be in the luxury position to say, these are the guys who really like the best. If you do that the right time, you might be in a position to take them out and and add it to your existing stable.
Right now also, I mean, even if you look at the lenders you're dealing with, those are also typically not the big boys. The real money sits in serving the big four banks and assisting them in addressing this need in this sector. So I can see how this is a model that could actually work for everybody.
Brendan Le Grange 29:53
And this is an opportunity for a typical data disruption type of event, right, because the credit bureaus were kind of tightly controlled marketplaces because it's a licence based thing. So you get a couple of licences in the market. Usually. In many countries licence base. You've then got to go out to the banks, get their data, get them to sign up on some agreements, to share the data with you. But all those rules in place, and you've got this hierarchy structured approach to getting the data in, whereas the FinTech can get in there, say, well, I'm making some assumption, making some guesses and I'm using big data to do this.
Oscar Koster 30:30
And if you get credit going, it's it's a lot of fuel economy to keep to stop, stop going. That's the basic principle we're trying to do.
Brendan Le Grange 30:41
Yeah, I think that's it. Thank you very much again for your time. And thank you for listening. This has been how to lend money to strangers, the podcast about lending strategies around the world and across the credit lifecycle. I was speaking to ask a cluster of big data scoring coms African division. If you'd like to speak to ask her. You can email him at Oscar dot Costa at Big Data scoring.com. That's Oscar with a C Koster with a K or look for his details in the notes below. And I'll see you again next Thursday.