Strategy meets data science when it comes to SME lending, with Frank Gerhard
“This is not a small change, and it doesn't affect isolated groups. This is not just an IT problem. This is not just a modelling problem. This is not something the business can solve on its own. This really requires board attention to actually bring together those teams and ensure that there is really cohesion across the organisation.” - Frank Gerhard, Solution Associate Partner and CTO at Risk Dynamics, part of McKinsey & Co.
Gone are the days when senior business leaders could leave the data crunching to a handful of propellor-hat-wearing quants in a dark back room, now there’s a seat at the Boardroom Table for data and analytics, or at least there should be. In this episode, Frank and I speak about the strategic elevation of analytics and how innovations from just a year or two ago are already mere table stakes, but in particular we talk about all of this in the context of SME lending where lenders are feeling them acutely.
I first found Frank via his co-authorship of this article on “How banks can reimagine lending to small and medium-size enterprises” but we also reference this article on “Designing next-generation credit-decisioning models” which is worth a read, and has more insights we had time to fully explore.
Frank is on LinkedIn and available via email, or for more insights from him and his colleagues, you can head over to https://www.riskdynamicsgroup.com/our-insights or https://www.mckinsey.com/
Frank also mentions Quantam Black, AI by McKinsey’s - their work is at https://www.mckinsey.com/capabilities/quantumblack/how-we-help-clients
He also speaks about Andrew Ng thoughts on great data versus big data, I’m not sure if this is exact article Frank meant, but it is the same philosophy: https://spectrum.ieee.org/andrew-ng-data-centric-ai
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.
Oh, and if you’re in need of more banking podcasts, you can find related content at https://blog.feedspot.com/banking_podcasts/
Regards,
Brendan
The full written transcript, with timestamps, is below:
Frank Gerhard 0:00
If you actually start at the board level: thinking about why do certain things not quite work? why are we losing market share? Why are we not growing as fast as we can?
I mean, once you actually get into the engine room, you open the door, very often you find data related topics, modelling related topics, infrastructure process topics, really at the heart of what's not working. And I think this is what drove the firm really to bring colleagues like myself and a lot others on board.
But also to bring on board entities like Risk Dynamics, really specialists on risk model development, risk model validation, but also entities like Quantum Black and others really focusing on data science and other engineering topics.
Brendan Le Grange 0:46
I imagine many students are like this, but late into my final year of university, I didn't really know what I wanted to do next. And I wasn't going to be picky. I just wanted that first job - so I sent my CV to every company in UCT's graduate recruitment handbook. There were maybe 30 of them with at least a tenuous link to my degree, from which I got 3 interviews.
There was the investment bank that had asked for a 100 word summary instead of the boring old resume. For them, I bought a coconut, emptied it and closed it back with a little bit of that silver chain that you get on the plug for your kitchen, holding in a few autobiographical notes. I called it, "Brendan in a nutshell".
And according to people who ended up working there, it was still being spoken about two years later... which gives us a sense of how much I must have blown that first interview, because I wasn't called back for a second one.
My next stop was with McKinsey and Company. I was nervous going in because, well because it was McKinsey and Company and they've always been hard to get into. Plus, I'd had to fly up to Johannesburg on my first adult trip alone.
I was even more shaken when the first question was about why my lowest grades were in maths. Now I've actually always been pretty good at maths. And I'd got a B, so arguably, this was more of a problem with their expectations then with my delivery...
But I did end up getting a wonderful job with Capital One. All's well that ends well, I suppose. And in the two decades that followed that fateful fork in the career road, I never bumped into McKinsey and Company in any credit risk capacity. To me, they were always firmly in the business strategy camp, representing some alternative future that another me had taken, never to be seen again.
Except in the last few years, I've started to see some articles coming out from McKinsey, suggesting I may have been wrong. Articles with quotes like "banks need to implement more automated credit decisioning models that can tap new data sources, understand customer behaviours more precisely, open up new segments, and react faster to changes in the business environment. And since there is no universal solution to credit lending models, banks and lenders must identify and design the lending process to align with their aspirations and business objectives". That sounds a lot like what we talk about on the show.
Welcome to How to Lend Money to Strangers with Brendan le Grange.
Frank Gerhard, CTO for Risk Practice at McKinsey and Company and co-author of the article 'How banks can reimagine lending to small and medium sized enterprises', which we'll jump into later.
Welcome to the show.
But for now, let's start with more of an introduction. I see you were a banker, like many of my audience, before you became a consultant so let's start there - what was your pre-McKinsey experience?
Frank Gerhard 3:43
Good beeing here, thanks for the invite.
Before McKinsey, I worked for 13 years on the trading floor at Barclays Capital, Credit Suisse, Unicredit - half the time in risk roles/ half the time in business roles. Financial econometrics was originally my academic career track, and I've really been following the combination of working on the business side as on the trading floor and working on economic statistics over the last 25 years or so.
That's really what makes me get up in the morning, and what's fun is really working on models, to be honest.
Brendan Le Grange 4:26
What made your decision to move from working for one client, as it were, to working in a world where you've got multiple clients and you're consulting.
Frank Gerhard 4:34
I think it started with my last big project at the last bank I work for, where we developed a large counterparty credit risk framework. And after being finished with that, the idea was to run the machine on a day by day basis. And to be honest, that's not really me.
So when at the time McKinsey came around, looking for a colleague to lead some of the work in advanced analytics and the risk practice, I was actually really fascinated by that. I mean, I was a bit surprised at the time, eight years ago, I wasn't aware of the analytics work McKinsey did, I only knew it for the strategy work. But after having talked to the colleagues there, it really sounded like a great opportunity to do at scale what I'm really interested in, working with clients on modelling risks.
Brendan Le Grange 5:15
Yeah, when I left university for my first degree, I actually interviewed with McKinsey, unsuccessfully, but instead then I joined Capital One and got into credit analytics. And I'd always seen that as a fork in my career, that had I got an offer from McKinsey, I would have done some strategy stuff, mergers, acquisitions, organisational redesigns, that sort of thing. But instead, I went into risk analytics. And for the first 15 years of my career, I never saw those paths converging. I didn't bump into McKinsey in projects.
But more recently, that's what I've started to see some of those articles of yours and your team's coming out some data visualisations in this space. Was that a change intentionally expanding into new arenas? Or was I, and were you, simply unaware of work they'd been doing in the background already?
Frank Gerhard 6:05
Just from what I observed, and from how I see it is that there was really a big push over the last 10 years to also work in the space of advanced analytics and to do that in conjunction with the McKinsey profile, and the McKinsey depth around strategy consultancy, right.
I mean, this is not just doing data science, but actually looking to bring together and harness really advanced analytics, modern methodologies to really bring forward business strategy on that side. And that is something specifically on the credit risk side, which I'm seeing more and more, where if you actually start at the board level thinking about why do certain things not quite work? Why why are we losing market share? Why are we not growing as fast as we can? I mean, once you actually get into the engine room, you open the door, very often you find data related topics, modelling related topics, infrastructure process topics are really at the heart of what's not working.
And I think this is what drove the firm really to bring colleagues like myself and a lot others on board, but also to bring on board entities like risk dynamics release specialist, I'm specifically work for on risk model development, risk model validation, but also entities like Quantum Black and others really focusing on data science and other engineering topics.
Brendan Le Grange 7:22
What is Risk Dynamics? And how do you fit into that broader ecosystem that's being built?
Frank Gerhard 7:27
Risk Dynamics got bought by the firm, six years ago, I believe, and at the time was a group of roughly 30 colleagues. Over the years, we've grown that to well over 200 now, doing work on risk model development and risk model validation.
So if you think, for instance, about credit underwriting - it's kind of a typical case. Think about the role models can take in credit risk underwriting to actually create, on the one hand, a really resilient risk management approach. But on the other hand, also business growth, which even works across phases like COVID, for instance, or the recent high inflation period we've moved into on that side, right?
I mean, models really sit very much at the heart of that. And this is kind of a typical area, we work on credit risk, for instance.
But also, if you think about topics around customer assistance, really working on what happens if let's say a loan did not quite work out and the customers need assistance on that, side topics around early warning system, if I think about nowcasting, and thinking about alternative scenarios, but also, once you move into, working with non financial institutions, working on financial risk topics, like understanding the financial implications of the supply chain, on the viability of the entity as a whole modelling these sorts of scenarios, I mean, that's all typical work.
And also, if you leave the financial risk area and go into non-financial risk, we're talking about topics like KYC, anti money laundering, are also heavily quantified topics nowadays, where very often the approach on the modelling side is also super helpful, to create, on the one hand, a good customer experience but on the other hand, also to make sure that the bottom line in terms of a safe KYC process and robust anti money laundering is actually met on that side.
So you really see the modelling approaches throughout, but we do focus in Risk Dynamics, as the name says, really on everything which is risk related, including model validation also, which if you think about the reliability of models, especially in phases, like we're seeing them at the moment, right?
Over the last three years, originally, I would have said COVID is a structural break. Okay, fine. One, two years afterwards, you have the next structural break, and then immediately the next structural break. So we seem to be going from one break to the next. So that's almost normal on that.
And that's in a way what we what we work on.
Brendan Le Grange 9:54
To see the likes of McKinsey working in the strategy, the 'this is how it fits in the business', I think it's really good news for everyone in the industry, because it means that you're not just sitting in a dark room in the corner, this now really is being discussed by the business as a business issue.
Frank Gerhard 10:11
The boardroom has actually discovered the importance of model quality, and even more importantly, of data quality on that side. And that is something which is a big change.
And I think that is driving a lot of the attention. You're observing on our end, but also really other players in the market, data and analytics are really central to what we would consider an edge in the market.
And if you think about time to market, and the ability to develop models quickly, that is turning into this strategic capacity. I mean, the reality is data and models have come out from this dark corner, or from the back rooms, and really taken a take a prominent role in the boardroom nowadays.
Brendan Le Grange 10:53
If we turn our attention to that article that I mentioned earlier, you start that piece by saying "if banks reimagined and modernise their business lending processes, they can take advantage of new opportunities with SMEs and capture more of the forecasted growth".
It's a simple but powerful introduction to the piece - why did you look at SME lending in particular?
Frank Gerhard 11:16
It's a good question. Actually, what we've seen with our clients is that for very established players, we really see eroding market shares of some of the established lenders there, we see growth lagging behind ambitions.
And we actually find that some of our clients are really looking at SME lending still as something which is part of the wholesale business, and in a way are applying either parts or entirely in some cases, the processes which are really designed to lend, let's say, three figure sums to large corporates. There are rights and justifications for doing this, but from our perspective, SME lending is really more of a mass lending business, you really need to think big on that side. It's not so much driven by papers getting to the credit committee, it's really more driven by models and data helping you to take decisions in a repeatable in a standardised way on that side.
Then again, on the other extreme, we see some of our clients really running the SME businesses out of retail, which again, ignores an important parts here: that actually SMEs require some special attention.
I mean, there is not one generic SME 'right way', if you think about shop retailers, tradesmen, doctors, or I have e-commerce merchants, right, I mean, all of these really need to have specific services, but also a specific way to navigate the application process, which is very different if you think about a doctor applying for a loan and if you think about an e-commerce merchant applying for a loan, you will find they have very different expectations.
And you will really want to be able to mirror those expectations in your processes. And if it's in the processes, it also needs to be in the models. Then you have to think about, well, what data actually allows me to do that. So there's a whole chain of topics around this. And that's really what drove us to think about SME more in depth.
Brendan Le Grange 13:11
And on that, growth is one of those words that you'll see in any number of headlines, but you really are talking about significant differences in growth. So you're talking about the example in your article of the SME lending bank that offers instant decisions in a fully digital format, they're growing at double the rate of the segment as a whole.
This is not maybe 10% faster than your peers, doubling the growth of the segments you're in, it really does show that there's significant gains to be had here. And I guess underlines the scale of the impact that can be made in the space.
Frank Gerhard 13:46
I fully agree with you on that side - the growth is very significant. #
If you put yourself into the shoes of let's say an ecommerce merchant. And if you think about how that person typically would run or how that customer typically runs their business, they're used to actually working on the screen, working on the computer, if you actually ask them for an in paper trail of their last five financial years on that side, right. First of all, they might not have five years of track record on that side, right. I mean, they might not have existed 2/ 3/ 4 years ago. And also it's very disruptive for the way they actually organise themselves.
I mean, it's just not the way they operate. Does that mean you want to economise on your risk taking? Or would you take more risk on that side?
No. It's not what we're saying. And it's interesting. You mentioned big data earlier. I think Andrew Ng, one of the proponents of machine learning, I mean, great, great educator in that space, said, well, we need to move a little bit from big data to great data. And I think that's also the case in this context, where if you actually put yourself into the shoes of a specific market segment, take that proverbial ecommerce merchant and you think about what's actually the great data, what is helpful to characterise to pinpoint the risk of that specific merchant versus versus his peers on that side, right?
Then, in a way, we're really heading in the right direction there because a risk model, or we're then actually enabling the business to offer a product and to offer a customer experience, which is actually irrelevant to the individual groups. The corollary to that, of course, is that we need to think about market segments. And we actually have probably more models, and really a higher importance on fast model development than we had in the past.
Brendan Le Grange 15:03
On that end, you published the article in May of this year. At that point, already, we were seeing the emergence of the situation where now we ahead Russia had invaded Ukraine, we were seeing inflation rising, but I think it's fair to say that it's a bit more gloomy. Now, when we're recording at the end of September than it was a few months ago. Do you think that the sort of environment of high costs high interest rates, persistent uncertainty, do you think that makes the push to digital more important?
Or do you think some lenders are going to step back and say, well, the balance of power has shifted towards me, I can erect some, some walls up there and the borrower is going to jump over them, there's going to be less competition, I'm more nervous of the market, maybe? Or do you think no, no, no, we're gonna still focus on making these changes and modernising our lending processes.
Frank Gerhard 16:21
As a good econometrics person, let me give you an observation on this. Since I would say the last six, six months, we've actually seen a massive step up, not so much on the collection side. But actually, on the credit underwriting side, I mean, the number of requests we're getting from clients to do this sort of deep analysis, especially in the SME space, but also in the retail space of their processes of their models of the analytics surrounding it to to make sure that it's really walking a fine line here.
I mean, on the one hand, we've seen in COVID, that just undifferentiatedly slamming on the brakes, is totally the wrong answer. Because what we've seen after the COVID lockdown, was actually a significant number of SMEs really growing very quickly. Just think ecommerce, I keep keep coming back to that example but it's by no means the only example.
But we've also seen other sectors or sub sectors in the SME space actually, really struggle - think restaurants, think hotels, think other tourism related sectors, it's really that sort of understanding, this ability to really triage and look at the market in an up to date way, not just using the data of the last five years, but actually bringing in current information being able to bring in, for instance, transaction data.
And if we're able to bring in this reliable view on the world, that growth is actually still there, and very important, but I would strongly recommend not just to continue in an undifferentiated way, what you've been doing over the last 10 years, characterised by low interest, low inflation, and so on and so forth, the environment is definitely changing. We see our clients adopt to that very quickly.
But adopting to it does not mean slamming on the brakes, it actually means getting more sophisticated in the analytic space, getting more sophisticated in terms of how can I assess the, let's say, affordability of a loan for a specific retail customer, for a specific SME customer on a case by case basis, in a scalable fashion. That is really where we see really a lot of interest and a lot of movement over the last six months.
Brendan Le Grange 18:33
It's all on that understanding that the past represents the future. And for the last decade, that's been true. Now, they've been so many shakeups that we know fundamentally that, well, we've got the last two years are COVID years, so they're strange in their own right, but also we know that the pre-COVID world no longer represents the current scenario, at least, you know, the next few years of the world.
So we've got all this confusion that calls into question that really important thing, that the past will represent the future.
But how do we adapt to that, you know? If we're just left on our own, everybody's going to run off in different directions, make different assumptions, tweak the model this way, shorten the data period that way, maybe use some other approaches that are more responsive to volatility, but they might be bumping into each other, they might be doing all sorts and maybe modelling is tightening the taps while acquisition or going into new market spaces.
You really do need that business thinking to say, well, why are we having to adapt, and to apply some of that old expert thinking on top of the models which I don't think everybody's going to be able to do if they just look at their own small aspect of the lending process. So I think yeah, that's probably where a team like yours is ideally placed to help an organisation think what changes, what adaptions they need to do, and yet not not overreact, not turn everything off, but just change how they how they see the world
Frank Gerhard 19:59
It's really bringing these these two dimensions together.
It's really not being aggressive on the risk taking side, but as a matter of fact being prudent on the risk taking side and really thinking about what do we need to do to actually make our processes on the one hand, more robust, model maintenance /model risk management has really come to the fore of over the last two to three years, right?
I mean, model risk management always used to be this sort of painful thing you have to do after you've developed a model, right? I mean, nowadays, model risk management is almost a strategic commodity on that side to ensure Brendan, as you put it, right down and that at a Board level, we really have the confidence that, okay, fine, at the moment, we have the right models, but we also have the right processes in place to find out when do we need to renew models? When do we need to reassess models, we actually have principles which work across structural breaks, and which allow us to assess and reassess.
And that this sort of assessment that reassessment model redevelopment is part of how we actually do business. You know, we talked about modelling and data coming out of the grey zone, we also see model risk management actually coming out of the out of that very unpleasant, boring area into something which as a board member you really want to rely on, right?
I mean, you want to rely on these people who tell you, yes, these models work. And they will stop working under those circumstances. And by the way, that is actually the list of models you need to focus on at the moment to make sure you can continue lending on that site. So there's, there's really a very big change in visibility and also in focus on these areas. Absolutely.
Brendan Le Grange 21:38
And what are the sorts of things that lenders should be thinking about if they want to check all the processes up to modern standards? Or, more likely, where they could be making changes to bring those up to the leading state?
Frank Gerhard 21:52
I think there is two things which are really essential to get started on this.
The first item is really attention from the boardroom. This, this is not a small change, and it doesn't affect isolated groups, right? This is not just an IT problem. This is not just a modelling problem. This is not something the business can solve on its own.
This really requires board attention to actually bring together those teams and ensure that there is really cohesion across the organisation.
The second one I would highlight, and that's, again, more a top down point, is really aligning the organisation behind the Northstar, behind one vision where you actually want to get to even if it might take two years/ four years to get there.
How do you imagine your organisation servicing SMEs, retail customers in one year, two years, three years time? What's that vision there?
And the reason why I'm emphasising those is because that then actually gives the sort of focus and attention and also positive energy to change things. And very often, where we see a need for change is really around topics, which feel a bit odd at first, which have to do, let's say with the end to end model development chain. Right?
I mean, something which is a very old topic, right? I mean, nobody would have thought five years ago to talk about the more end to end model development chain at a board level. We're talking at the moment to CROs to Chief Lending Officers about exactly that, right. I mean, how do you get models developed quickly, because these models then enable customer friendly processes, customer friendly processes really allowed to, to drive really the business development on that side.
And it's really that sort of pursuing these topics into the engine room, which really allow us then to dig into this. There's also a couple of other corollaries to that one of those is also the infrastructure of taking credit decisions. If it takes you six months or even more to actually get a new decision model into the system. This is not going to work in this environment, you really need to aim for a very quick ability to turn around, to actually develop the models.
And once they were developed to actually release them into production, once they're released in production, you really want to have a very close eye on the quality of these models and the performance of these models. That is very much the end to end chain on that end.
And then finally, we talked about it earlier, great data.
It's not always big data, in some cases, transaction data it is, but very often it's really great data! It's really thinking about what gives you a line of sight on the information set, you really want to measure what's the what's your customer group you really want to look into here. And what's the information, the data, which really helps you to understand the opposition in the peer group very often, this is really an exercise which is not fully done, and which which really unlocks some of these other steps on that end.
Brendan Le Grange 24:42
In terms of volatility, are they different technologies or the different approaches that you're seeing now to deal with that sort of quick turnaround? How are you helping customers to be in a position where it is possible to roll out a new scorecard that is safe and monitored within a matter of months?
Frank Gerhard 25:00
Yeah, you know, we started actually looking at our own processes. I mean, if I think about the 200 analysts we've got at Risk Dynamics, we actually started on that ourselves where in some cases, we said, okay, fine, now we're doing the same model for the third time, fourth time, tenth time and we felt that we just weren't quick enough on these things, given we've done it already.
So what we did is we looked around, and we found actually that software engineering is a great area to think about exactly those topics.
Versioning code is trivial, right? Versioning data, that's a different story, versioning code data, the documentation, pipelines, side by side, being able to reconstruct that over time trying out new things in different versions, things. That is a different story.
And that is really where we're going at the moment. And we're we have been working with a number of clients over the last 12 months is really taking these lessons from software engineering, but let's face it, statistics person or econometrics person, like me is not a software engineer. And I'll never be one. But I can still use some of the tools if they're appropriately set up and appropriately tailored on that.
That is one of the contributions on that side is really making sure that we set up the sort of a work environment of a model developer, very much so that it feeds into the validation, it feeds into the production release. And when we think about the production release itself, that we may be, get away from this notion that every production release needs a full squad of engineers to actually implement the new model into production, right? I mean, maybe we're doing that in programming all day long, right? I mean, we parameterize things. And releasing a new set of parameters is a totally different story, compared to releasing a full new model built on that site.
For some reason, we still feel very often with these productive releases that we need to do this full rebuild when all we're intending to do is just a modification for a sub segment or an adjustment. And it just takes too long. That is really very much on the let's say engineering side of things, or on the process side of things on how we do things. The other very helpful piece of information we've seen over the last few years is really transaction data.
I mean, the use of transaction data, and I would argue is is becoming table stakes for what we're doing here. It used to be a competitive edge or a nice to have framework, but it really turns more and more into a must have for some types of land to really have that sort of visibility, either on the transactions, or maybe on the sales flow of some of the SMEs as we might get them from some of the ecommerce platforms are really some up to date signal on the performance side.
Because otherwise, we might be struggling to actually get this sort of up-to-date confidence that this is working the correct way. I think that's probably the two three most important ingredients I would see at the moment. B
ut maybe just just one more aspect is also machine learning, which I think has been, if I go back 5-10 years, has really been hyped as the solution to a lot of things. It's it's quite interesting to see that, yes, machine learning methods have certainly a place in the toolbox, but the emphasis is on a place. And it's by no means the most prominent one, especially at the moment.
But what we're also seeing is that for machine learning, there is a great emphasis nowadays really on topics like addressing potential bias. I mean, I'm sure you've talked in your broadcasts before about some some of the notable examples. In that case, explainability of machine learning is quite essential. And quite recently in research, we're actually seeing this cross pollination between econometrics statistics and machine learning, really producing some really amazing methods, which over time will find their way into BAU applications.
But we see a little bit of a, let's say, you're stepping back off the let's say, conventional machine learning for the benefit of a 'okay, fine, so how does this methodology actually help us in a specific context here and how can it actually contribute to the overall effort rather than being the panacea for everything on that end?
Brendan Le Grange 29:12
Yeah. And I think that's probably also some of that maturation of the people involved in the space. And we now do have analytics leaders, business leaders who've worked on machine learning, and we all understand machine learning a bit better from the early days when we're like, well, is we've seen it in the movies, we throw data at it, or we do what it tells us to do to this appreciation of Okay, now we've got a better understanding of the boundaries, what it's good at what is not so good at what the regulator likes, what the regulator doesn't like, what our boards want.
And now it's a bit of more of a tool that we can in this business sense sense use, which is more powerful overall, because it's a bit more under control.
Frank, if anyone listening would like to work with you and the risk dynamics team or just read up some of that research that you publish way is a good place for them to go to learn more
Frank Gerhard 30:00
Look at our website Risk Dynamics (https://www.riskdynamicsgroup.com/). So if you just Google that one, you'll you'll get there Risk Dynamics Group, very straightforward. If you want to get in touch in person. I mean, obviously, my LinkedIn profile is there, otherwise, you could drop me an email frank_gerhard@mckinsey.com
Brendan Le Grange 30:14
Great, Frank, thank you so much. I will put those in the show notes as well. Thank you so much for making the time today, I think really interesting strides happening at a really interesting time for modelling as we really do have to grapple for the first time with how do we embrace all these new technologies, and a world where last year is very relevant, but not as relevant as last month and we turn that to the future. So great time to be talking to you and thank you for your input.
Frank Gerhard 30:39
Thanks for the great discussion. Really enjoyed it. Thank you.