Raymond Anderson gives us a history of risk assessment

How to Lend Money to Strangers will be a far-ranging show, in time hearing from lenders big and small and in markets across the developed and developing world. However, before we embark on that adventure, I wanted to establish a foundation, to create some context.

So for my first episode, I sat down with Raymond Anderson to talk about the history of risk assessment: how and where lenders first came to see value in sharing their data, how innovators then created and improved scoring models based on that data, and how the drive for improved scoring techniques continues to evolve today.

Raymond Anderson is a credit scoring specialist, consulting to clients around the world as proprietor of Rayan Risk Analytics and as an advisor within the IFC. He is also a globally recognised speaker and author on the subject, with two books published by Oxford University Press: The Credit Scoring Toolkit is available from all the usual channels, while Credit Intelligence and Modelling will be released later this year — we discuss those, too.

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

You can read the full transcript with timestamps here:

Raymond Anderson 0:00

So many people that I speak to, are still heavily focused on banks. But where did this start? It didn't start in banks. It started in trade credit. And it started in store credit. The banks were extremely laid off. And as far as I'm concerned, they're still later adopters. It goes a lot slower, you're just scared, you're gonna keep on having the steam engine instead of a Ferrari.

Brendan Le Grange 0:53

Welcome to How to Lend Money to Strangers, the podcast about lending strategies from around the world, because let's face it, not every show can be about true crime. I'm your host, Brendan Le Grange.

I have been working in consumer credit risk for the last 20 years within lenders and consultancies, delivering projects across the credit lifecycle, and in over a dozen countries, from Africa, to Europe to Asia. And for most of that time, my job was to translate analytical potential into business-affecting stories. That's what I want to bring to this podcast.

I'll be speaking to a range of technical experts, business owners, and those who have a bird's eye view of the local lending market to hear how they think about solving the problems inherent in lending money to strangers. These stories will vary greatly from Bank to FinTech, from developing market to developed, from startup to incumbent to regulator. But if there's one constant, it is that these days just about every aspect of lending touches on data, and usually via some form of statistical model or scorecard.

So for my very first episode, I wanted to bring in one of the grand masters of that art, you heard him right at the start reminding me that lending is not within the sole purview of banks. It isn't today, and it never has been. Banks and their concerns must be a keystone of any discussion on lending strategy, of course, but many of the tools we use today were not innovated inside of a bank. And many of the tools we will use in the future are similarly likely to come from the outside.

Raymond Anderson literally wrote the textbook on credit scoring, two of them in fact: The Credit Scoring Toolkit and Credit Intelligence and Modelling. He is a globally requested speaker on topics of risk modelling, and something of a star in China. And I was in awe of the breadth of his knowledge on the subject. When I had the pleasure of speaking to him for this episode, we spent a good amount of time talking about how the industry got to where it is, before taking a closer look at some of the subjects that will be recurring themes in the show.

So I think in terms of kind of kicking it off, whether it be really great to have some contextualising of the history of predictive modelling and credit scoring in general, like where does it come from? How did we get to where we are now? Where do we think we're going?

Raymond Anderson 3:10

Alright, I would put it down more in terms of a history of risk assessment, of ratings and scoring. Allow me to ramble for a bit, the extension of credit has been going on for practically as long as man's been sentient, or there's been sort of some type of back and forth, whether it be in a barter economy or a money based economy. But initially, it was always based upon personal relationships. And if not personal relationships, potential sanctions, collateral, most of that relied upon the knowledge of an individual, the person who held the purse strings. The real thing started when you start getting into the instances, like the Middle Ages in Europe, where you have merchant banking, branch offices, and potentially amongst the Jewish community, where they were doing the money lending, where it was rules that were learned and passed on to employees, or to sons and daughters and family, sort of other people who are picking up the trade of moneylending.

And even then, it was based upon getting information, much of it was down to the extent of having agents and foreign courts. The Italian moneylenders had agents in the English court in the 15th century, who were reporting back because the king at the time was going on all sorts of escapades into France and sort of like warring all over the place and they were worried about the money they spent - it actually brought two Itallian family banks, it ruined them, it bankrupted them.

The one thing that came about with the Industrial Revolution was collaboration between lenders. The first real effort at collaboration was in London, the Georgian tea houses in the 18th century. And most of it was just informal at first blenders making coffee or tea houses. And in 1776, there was a society formed called the London Society for the Protection of trade against swindlers and sharpers were swindlers and sharpers were frauds and cheats. And this was the era when no distinction was made between fraud and credit. So if you were guilty of either, there was a potential for ending up in the debtors prison. What no one really knows what happened to that society. But it laid the groundwork for most of what happened in the United Kingdom thereafter, where you had a multitude of trade protection or guardian societies that developed in various cities across the United Kingdom, or sometimes they were collaboration between lenders, sometimes there were established by chambers of commerce. One of them that started off being seven tailors serving the carriage trade in an era where you had a whole bunch of dandies and dandezets that were going around with fancy clothes, living off of their appearances and status within society and not paying their bills. That still often happens today... some things never change.

But this whole thing, this mutual society type of arrangement, was not peculiar to the United Kingdom or Europe, because it also applied to a lot of consumer lending in the United States. But in terms of the trade credits in the United States, it went on a for-profit basis, where you had enterprising individuals that would set things up. We'll cover that in a moment. The other type of setup was government initiatives to establish credit bureaus. But the one thing you'll find with the government initiatives is that it's often the government will find them doesn't have the capacity to run it, and it will either offload it to a private concern or engage with a private concern to provide the technical expertise to run it. Over the past years, there have been a lot of new bureaus established and ever smaller countries, much of it has been an initiative on the part of the International Finance Corporation of the World Bank, trying to speed the development of emerging countries.

But get into it the more recent time sort of the same - now more recent, I'm talking here, early 1800s - Barings Brothers was using a Mr. Ward on the east coast of the USA as an agent, a spy gathering information on the ground. It was a very expensive way of gathering information. In the 1830s, there were a couple of crowds that started experimenting with providing a service where it wasn't just one person providing information to one company. It was a business enterprise that was going to collect information and share it with a subscriber base type setup. And it was mostly started out of people who are involved in wholesaling. You had imports coming from England, or the factories on the eastern seaboard of the states who were serving the country trade. The people would sort of arrive on mass, the traders would arrive on mass in New York and Boston and sort of wandered by and some of these guys developed dossiers of information for first and foremost was Lewis Tappen. He was a sock importer and wholesaler in New York. And Lewis had a fairly failed enterprise elsewhere, but he was doing the back office was interviewing clients and developed a lot of information over time and people started coming to him and asking him for date information of these dossiers and he would tell him, when he said, Hey, this is a business opportunity and became known as the mercantile agency in the United States and other mercantile agency best that I can tell was not company, but a network of agencies they went by various names depending upon who the proprietor was, whether it be in St. Louis or Boston or Philadelphia or Baltimore or wherever it was set up. They extended to such an extent that at one stage they were considered practically the largest employer in the United States. They say largest employer, not everybody was an employee, though they worked through what they call credit reporters. And the credit reporters tended to be lawyers, attorneys respected members of the community in various towns across the United States, three of which went on to become American presidents. Actually, not four went on to become American president. Abraham Lincoln was one of them. Now, they kind of dominated the kind of American trade.

John Bradstreet, he started up John Bradstreet & Sons for the country trade blah, di, blah but it was only a few years later. And it was really Bradstreet, who was the first to come up with a credit rating concept of the credit rating, and about 1857 of the mercantile agency eventually passed on to Robert Dunn, and Dunn was a bit slow to catch on, he was expanding, he did a hell of a lot of stuff. But Bradstreet came up with ratings first and Dunn came out perhaps a couple of years later. And these two were competing sort of forever and a day. It was only in the early 1900s, with |John Moody, he got the bright idea of saying, oh, yeah, well, we're doing all this and perhaps we could apply the same idea to traded bonds. During that time, all was judgmental. There were no pointing systems. But during the 1930s, you had a couple of people who presented papers where they'd come up with pointing systems that they tried to apply within banks. You had one of the home lending organisations during the Depression. They came up with a model, but it wasn't a pointing type system. Spiegel was a furniture company who sold on mail order, they came up with a pointing model goods were sent out without prepayment. They had the slogan, we trust everybody. And the level of morality within the United States, they weren't far wrong. They could trust everybody part of the American way was to pay your bills.

But the first attempt at a proper statistical model was David Durham, but it was purely an academic exercise. Effectively, what he was trying to do was to assess the validity of the decisions that were being made by companies selling motor vehicles. His claim to fame was really sampling. I think he got a sample of something like 7000 car loan applications, went through them, sample them down and came up with a model and said, Well, this model works. But he concluded that it would never be good enough to use practically. And during the 1950s, I believe Sears, I cannot find any proof of it, it's kind of more anecdotal, but I believe Sears also used scorecards of some sort.

The first real attempt now this is kind of the now where you start getting into what people commonly accept as the proper origins, credit scoring was Bill Fair and Earl Isaac, the origins of FICO. They were in California, they started, I think they were at Stanford University. This was in the era when computers were becoming big, and everybody was trying to automate as much as possible back-office processes. And the first thing was to come up with a billing system for a credit card called Carte Blanche. And they got the idea of saying, okay, we're advancing statistics, we're advancing and computer technology, can we not come up with some form model that's going to be used to assess credit. They had a mailshot, they sent the mailshot out to 50 companies, and they only had one response to the American Investments company. It was also a small loan lender based in St. Louis. But it had branches operating out across the United States, but they didn't initially for the St. Louis office, and it was successful. And eventually, I think they had something like nine different scorecards for different parts of the United States that were in play. And this was really Fair Isaac's proof of concept. The thing about it at the time, it wasn't just the statistics, it was also just a very tedious process, because so many of the applications were on paper. It was mass of data capture exercise, you would have to sort of set up banks have sort of captured terminals, punch cards, all that sort of thing. And it was a factory type process trying to come up with a single scorecard. But they basically captured the market sort of up until the mid 1970s, a period of 15 years.

In the 1990s. You had others coming up. One of the first was MDS or Management Decision Systems that were founded by John Kauffman and Gary Chandler. And they saw that FICO wasn't paying much attention to the information provided by the credit bureau. And they said, well, shouldn't we be trying to integrate this information in somehow they went out and they approached the same company that FICO had started off with that being an American Investments, and said, can we develop a scorecard for you, and before long, they were offering their services to the credit bureau themselves, MDS eventually got bought by Experian - actually, at the time, that was Consumer Credit Nottingham - and they bought MDS, at least partially, I believe, to get access to these scoring methodologies that they were using which was linear probability. That was Experian's start into the scoring game. And another crowd that emerged also for the same reason was Scorex, and Jean Michelle Trous, he was the European representative of FICO, perhaps a 10th, employee of FICO. And he grew Scorex into quite a significant concern. Unfortunately, he was killed in a plane crash on his honeymoon. The common feature here was that, like banks were slow to be on the take up of the scoring methodologies, FICO was slow to see the value of bureau information. And for that matter, the credit bureau saw FICO as a competitor, they didn't see FICO as somebody that they could collaborate with. And yet nowadays, a FICO score is synonymous with a bureau score.

Brendan Le Grange 17:29

Yes, I was gonna say, because people working in the industry will know FICO for many of its products. But if you asked a person in particular in the States, what is FICO do they're going to say 'the bureau score' so it's interesting to hear, they're actually pretty late to that game.

Raymond Anderson 17:42

Very much so. 1995, with credit scores being required for Home Loans within the United States, you have two, two legislative aspects to it. The one part forced some of the practices that we would associate spying like newspaper clipping, interviewing neighbours, bar gossip, that was pretty much banned. That forced the credit bureau to start saying, okay, we need to get more concrete information. And basically, it said, okay, you cannot use your judgmental biases discriminate against anybody. So it really forced people towards the statistical models, which could be used to support decisions. So if you can support the decision by numbers, you are seemingly free of bias, I mean, that there is an argument there, a valid argument relating to the disparate impact. But in any event, you had all of these forces that were shifting the trend towards empirical analysis of potential risks and applying number crunching as part of the process so much initially was for the small loan lenders, retailers, credit card companies, not the banks, the banks say we're kind of have the view with door lock, we've got the capabilities. And as a result, a lot of the banks were reluctant to partake, that applies mostly to Britain and the British colonies, where you have the four major banks, five major banks, it was more fragmented in the States. So the, the American banks are more open to it even if they were probably also a bit slow.

Brendan Le Grange 19:34

When you spoke earlier about, you know, those first scorecards, the FICO model on the one hand, focusing on maybe modelling technique, and other competitors coming through with a broader data approach... it sounds a bit like its history repeating itself. There's a lot of FinTech happening in alternative data, in machine learning, and in new modelling techniques. So we seem to be in this position where you've got again two roots and at some stage in the future, once, you know, once it's a little more mature in terms of AI and machine learning, we may then see those coming together again?

Raymond Anderson 20:07

you mentioned about machine learning, I see machine learning as being something that can provide value. But you need to sort of use it with some scepticism. For the most part, you're going to use machine learning where the materiality is low. Banks are not great adopters. Small loan lenders would be more open to it, the fintechs, or you've got a cell phone company whose extending credit - those are relatively low materiality applications. When you start getting into vehicle finance and home loans, and especially where there is a significant requirement for transparency, where they say, 'listen, so you can use this stuff but you have to know how it works'. You can have the best possible prediction, but if you don't understand it, then it gets marked down.

So if I were to advise anybody, I would say Listen, if you want to go and try out the machine learning, always have a backup, be ready to show out the machine learning model or whatever reason, or potentially use them in conjunction with each other. But there is more to be gained out of data, than the greater sophistication of techniques. There is a flat maximum no matter what technique to apply to it. And one of the most obvious, which is happening in Europe, is the open banking, primarily related to transactional data that banks will have a huge amount of, and which is given banks a competitive advantage over smaller players. We're only just at the start of it in South Africa, but one could imagine the whole Open Banking thing being not a separate type of credit bureau but being taken over by the credit bureau. The purpose of it was to open up the lending market to smaller lenders and to FinTech. The goal was a base of transactional data that could be used in particular to support lending to those people that don't perhaps have that credit history. I developed a model based on some transactional data in Kenya, and all I had was the minimum balance the maximum balance and the credit turnover for a transaction account or current accounts. And what came into the model was the month on month volatility in those figures, I just use the coefficient of variation for six months. And for both of those values it showed up in the model. And they don't didn't have a strong Bureau so I'm not able to say how it wouldn't have worked at the bureau information without but suffice to say that was sufficient to feature in the model and not insignificantly.

Brendan Le Grange 23:11

And so the UK bureau environment already includes turnover on current accounts. So there's already a pretty solid understanding of affordability. And even in that space, putting the open banking data in gives a noticeable improvements in in risk modelling. In terms of credit risk when I was in the bank, I don't think we were using 'actually what transactions that you spend on what merchant category codes?'. A few fraud or fraud adjacent things like okay, spending casinos we might look at differently or cash withdrawals, but I think, this Open Banking, even with the big banks, I think it's opening their eyes,

Raymond Anderson 23:43

When you're talking about that. That's actually one area where I would see that the machine learning models most certainly would add value where you're willing to try to get that little bit of extra lift and find out where you're missing the boat or being locked out of a party then bring that in. But I would see that as being best as an overlay.

Brendan Le Grange 24:02

Obviously, we spoken at the moment about kind of a historic emergence of sharing data building scores. But that's not a process that's finished... credit bureaus. national level scores are still being rolled out. In my work. Recently, I was involved in the first rollout of scores in fairly developed markets like Thailand and Malaysia. You've been working with the IFC closer to the coalface and so it would be really interesting to hear how these same steps in the process are happening in the modern context, as well as the hard won lessons of rolling some of these things out but in a developing market.

Raymond Anderson 24:38

I'd worked for Standard Bank for 34 years of that, about 19 we're in credit scoring in various capacities. In 2014, I received a phone call out of the blue from someone in the International Finance Corporation who asked if I'd be willing to assist on a project in Lebanon. And at first I thought it was a joke. I didn't believe this, like, 'who is this guy?' And I listened to them. And I said, well, I'm employed by Standard Bank, but that shouldn't be a problem, we have lots of guys that are working for companies and they are allowed to do these projects, because they tend to bring knowledge back to the workplace. About six months later, I get another message saying the project is on. And I committed to it. And then the my employer, or at least my boss, knew about it. But the moment one started going through the whole process, they said, sorry, you cannot accept paid employment from somebody else while being employed with us. And I thought, after so many years, I think I need a change. So I went off to Lebanon, we sat around in a room for days, with credit experts trying to say, okay, we want to consider this, we want to do that, we want to consider the other things and we came up with a model. Now, one problem I always find with these things, though, is that you develop a model, you go away, and you never get any feedback. And often those models will never be implemented. Or if it anything goes wrong, you never hear about it. To the best of my knowledge, this was implemented and they were satisfied with it. Another one that I was involved in, and also rather different, was in Pakistan. But I was in a hotel that had a machine gun emplacements street facing the machine gun emplacement over the parking lot. This was in Karachi. If that doesn't make you a little bit nervous, then hmmm, it's kind of credit scoring on the far side.

Brendan Le Grange 26:45

Yeah, I have stayed at a few hotels, where there's x-ray machines at the door and maybe some of those guys with mirrors on sticks looking for bombs under cars, but never an actual machine gun turret

Raymond Anderson 26:55

And both manned. But in any event, in another instance... Now this wasn't the IFC, it was through the German Reconstruction and Development Agency, and it was in Myanmar, a lovely country, if it didn't have the political problems. Ultimately, the model was implemented with this simple rules based model and several months later, I received a phone call or a message saying they had seen that the scores were somehow corresponding to their decisions, that it was moving in the right direction, in the right fashion. Another one was Romania, that was one where I was presented with a very small amount of data, I only had 267 defaults. And then Kenya, the one thing I liked about the Kenyans, it was a very much a can do attitude, it's like let's get in there, let's do it. They want to move into the digital lending. Really, all I had to work with was some behavioural information on the behavioural information wasn't necessarily in the best form, and at the end of it having an application model with no applications, just purely based upon the behavioural data now got all of this done. And they were all very gung ho, but they didn't have the capacity to implement, to link in to their system. I think that happens in a lot of instances when you get into the consulting end, where there isn't follow through. We had a consultant for many years with Standard Bank, his name was Jes Fremantle and it was possibly a decade, if not a bit more than he was guiding us.

Brendan Le Grange 28:42

You're not just seeing this working in peak established markets. You're seeing this approach working in all conditions.

Raymond Anderson 28:48

Actually, one of the most interesting models that I saw was used for poverty scoring, where the guy called Mark Shriner, and he kind of had this idea where you've got all of this information from census, but you wanting to go out into the field, and you have these definitions of poverty, but it actually takes a lot to collect that information. But you've had a lot of field workers that are going out, and they want to do a quick and easy assessment of should a person qualify for some type of social assistance. So he had all of this data to be able to define who was poor who was not poor a lot of stuff around it. But he chose for his predictors, it wasn't just what was most predictive, he chose the variables that were the most easily observable or as ascertainable for field agents. That's something that we perhaps forget, we tend to sort of have data arrive at our desks and think it's all all free. So in any event, the field agents that go out, they look at her face, and they say, how many inhabitants are there? What's the roof look like? Is there a television? Is there bicycle? and so on and so on and so on the answer to the scorecard may say, Okay, this guy's he would qualify for assistance or not, it was perhaps one of the most novel applications that I found for applying these types of methodology.

Brendan Le Grange 30:15

In fact, one of the episodes that's coming up after yours is with some guys that are doing big data scoring in Africa, well in developing markets around the world, trying to scale that sort of approach. So they take NASA satellite photos and look at how much light is there at night to see how close you are to a school or to a train station, and incorporating that into their scores for thin file and new to credit customers.

Clearly you're keeping yourself very busy with your consulting work, your work with the IFC, but on top of that, you're also a writer. So maybe we can finish by taking a closer look at that. The Credit Scoring Toolkit has been out for over a decade now and established something of a following. In fact, some of my colleagues in Hong Kong were a little bit starstruck when they heard that I'd once worked with you, thanks to the rather popular Mandarin translation of that book. But you also have a new book coming out this year. So how did you move from scoring practitioner into textbook author? And what can we look forward to in your new book when it's released later this year?

Raymond Anderson 31:26

It started off being a little guide that was supposed to be written for the South African Institute of Bankers, as course material. The SAIB never 'got it'. In total, it was over three years of work. Then I met up with someone from Oxford University Press in Edinburgh. And they said please send us the manuscript. The end of 2006, it was finished, the manuscript submitted, and the whole thing was ready for August for the Edinburgh conference, a credit scoring and control conference, which is a bi annual conference and something I would recommend to anybody. But in any event, this next one's going to be virtual, it's coming up and I've got a paper that I'm looking to present at it. That said, it got out there. And in short order I was hearing, oh yes, somebody saw it in Uruguay, and someone saw it somewhere in Siberia, and someone saw it... So I realised that it was getting out all over the place. Now one of the things with these books is that they don't have very large readerships set of academic textbooks that you're probably be lucky if there are 400 eventually sold. Well, this one, I think it's somewhere towards 2000, which is not exactly Harry Potter, unfortunately.

Brendan Le Grange 32:55

No, but I've got two books out as well. So I understand those numbers.

Raymond Anderson 33:01

The thing is that it breadth, just sort of how widely it is read and how much it is referenced. There aren't exactly a lot of other books that are written by practitioners that go into any level of detail. There's really myself, Naeem Siddiqi, Bart Baesens, he gets more into the academic aspects of it. Another one is Steven Finlay. And all of them are good books. If anybody would buy my book, I'd say also read the others. Elizabeth Mays, her stuff is quite practical, McNab and Wynn also very accessible, different with mine, I think is the writing style tends to be quite conversational. An ex-colleague said, 'jeez, it's just like listening to you talk'. It's that accessibility level of detail that grabs people, it is a challenge to make a dry topic interesting.

In any event, I wanted to reach a broader audience. So I went back to Oxford University Press and it's currently in the works, it was accepted. This one is called Credit Intelligence and Modelling: Many Paths Through the Forest of Credit Ratings and Scoring. The credit intelligence aspects is tied in with the identification of data, the extraction of data, the analysis of data and the dissemination of information after the analysis to people that need it. I went into a lot of depth on the history of intelligence agencies, the history of credit before getting into things like business processes, the types of statistical techniques that are used. Much of it has an overlap with the first book, but greater detail. And as I dug into it, I found myself looking into aspects of statistical theory that, you know, one doesn't necessarily stumble across readily. So although it might have started as that the content is barely recognisable. For every page that I write, I probably have to read 1000. And a lot of what's presented is a synthesis of stuff from other sources. A criticism that was raised was that 60 to 70% of this, you can find other places. Yeah, but do you want to have to read everything else that I've read?

Brendan Le Grange 35:43

Everything these days, you can read somewhere else, but yeah, what of it is valuable?

Raymond Anderson 35:47

Very, very true. Very true. And I would like to think that I've come up with a couple of approaches that are innovative and that other people employ, approaches that I've used to good effect.

Brendan Le Grange 36:03

Thank you, Raymond. It's been a pleasure and an education. Now, we purposefully kept the level of today's discussion very high, focusing on foundational concepts and building context. But in the corporate world, were seldom given time to do that before were asked to demonstrate real business benefits. So in next week's episode, I address that Graham Whitely of Quid Pro Consulting puts his money where his mouth is when it comes to that, and has been known to structure his contracts so that no fees will be charged, unless certain revenue targets are met. I asked him how he links his lending strategies to real and measurable revenue gains. And we talk a little bit about champion challenger, and how to structure those tests. Join me next Thursday for that episode.

How to Lend Money to Strangers as a podcast about lending strategies around the world, and across the credit lifecycle. You'll find all episodes on Spotify, the Apple Podcast Player, or wherever you're listening to this one.

Raymond Anderson 37:29

Interestingly, so South Africa had the first carlone scorecard in 1978. But the second generation scorecard in 1983 was programmed into a calculator, an HP 41 c calculator if I remember correctly. I programmed it. I struggled to get scorecard and I had two bytes remaining on this calculator, so it was tight...

Previous
Previous

Graham Whitley is turning scores into revenue