 Okay, welcome back to this second session of the ECB annual research conference. It is my pleasure to introduce the next speaker, who will be Catherine Casamata from Toulouse School of Economics, who is going to present a paper on lending and monitoring big tech versus banks. This is joint work with Mathieu Vuvard and Rui Xiong, both of them from Toulouse as well. So Catherine, the floor is yours. The paper will be discussed later on by David Martinez-Miera from Universidad Carlos III de Madrid. So Catherine, you have 25 minutes for your presentation. Thank you very much Oscar for the introduction. Thank you for inviting me and inviting this paper to be presented here. So we had a macro session and I have to warn you, we're going to turn to micro models. Okay, so over the last decade, we've seen a rise in new players in the credit market. And some of these new players are pure fintech, pure financial players like peer-to-peer lending platforms. But we've also seen the rise of so-called big tech, that is players which have another activity like e-commerce platform and which decide to provide financial services and in particular credit to the users of their platforms and in particular credit to merchants. And if you look at the evolution of lending volumes of these new players, you can see that it increases over time. But you can see also that in the recent years, the volume of credit provided by big tech has exceeded that of pure financial players. And of course the question that we are going to try and answer here is what can explain the entry and the growth of big tech credit in the credit markets. Now in the literature you can find several reasons why these big tech have started competing in that sector. Most of them rely on the idea that, okay, big tech, they have a competitive advantage vis-à-vis banks, either because they have access to more data or because they have better technologies to exploit this data, which leads to them being more efficient at either screening borrowers or maybe at capturing cash flows, at sizing cash flows and preventing merchants from diverting those cash flows. And that's the spirit of some recent papers in that domain. Now honestly, whether big techs are more efficient than banks have access to more data is subject to discussion, probably to empirical analysis. You can find arguments in both directions. Now some people also mentioned that maybe big tech face less regulation, which makes it easier for them. Now what we want here is to depart from these reasons, which might be reasonable or not. And we want to focus on another set of reasons. We want to understand, besides efficiency motives, what could be the incentives for big techs to enter the credit market? And we are going to rely on a very basic reasoning. If you are a pure lender, a pure financial player, of course your decision to grant a loan will depend on the cash flows that you obtain from that loan. Now if you are an e-commerce platform and you decide to grant a loan, of course you will take into account the cash flow that you can obtain from that loan, but you will also take into account the activity that it will generate on your other business, the activity that will generate on your platform. For instance, you might benefit from having more merchants directly, and indirectly having more merchants could increase the number of clients that decide to go and trade on your platform. You could have these network effects. So that might change your incentives to provide credit. Now let me immediately put a caveat to that reason. The platform already has a lot of instruments to monitor the activity on the platform. It can set transaction fees, access prices both to merchants and to customers. So if we want to understand why big techs enter the credit market, we need to understand why it matters to jointly decide on the pricing on the platform and on the offering of credit to merchants. And this is why our objective is to analyze this joint decision of the platform. So the key ingredients and the takeaways, key ingredients will have some merchants for which financing matters. There will be financial constraints. We'll have a platform as well as banks with some ability to monitor merchants and to alleviate financial frictions, but we will not assume that the platform is better at monitoring or at reducing financial frictions than the banking system. We'll consider all cases. And another important assumption is that the platform will have market power in its core business. What are the main results that we derived in that paper? I think the first one is to say that the platform will enter the credit market even if it's less efficient than banks at providing credit. And that will be more precise in what sense I mean more efficient. And the second result is that the platform will not try to cover the whole credit market. There will be an endogenous segmentation with the platform targeting the more financially constrained merchants only. So in doing so, if you want, the platform will be a complement to bank credit, but unfortunately it will also substitute to some of the bank credit. So the entry of platform into the credit market will both increase financial inclusion if you want, but it will also make them compete with banks. And last, we studied initially a case in which the platform is in a monopoly, and if there is more competition on the e-commerce business, we can show that this will induce less platform credit, but maybe overall more financial inclusion. So in that paper, maybe what is the original approach is to try and combine two literatures, one from IO, especially from two-sided markets. I gave you only the first seminal papers in that literature, and the other one from corporate finance. Let me skip that and go to the model. So it's a simple model, and I'm going to make it even more simple by showing you pictures. So the basic two-sided market model of Rocher and Tyrol. This whole paper is a tribute to Jean Tyrol, and I'm mixing two models by Jean. So you have a platform, which is in a monopoly, and that platform helps merchants and consumers trade. Now, when they make a trade through the platform, the platform incurs a cost, too. And why do they make trades on this platform? Well, each merchant has some valuation for trading with a consumer, VM, for each trade. And each consumer has some valuation for trading with the merchant, VC, per trade. And of course, we have heterogeneous merchants and consumers. They differ in their valuation for the trade. And the platform will charge a price to merchants for using the platform for each trade and a price to consumers for using the platform as well. Now, from that basic model, we changed two things. So for simplicity, we shut down the heterogeneity in valuation across merchants. They would have another source of heterogeneity, but we'll introduce the fact that there are financial constraints. So the heterogeneity across merchants works as follows. They have different initial wealth. So they are indexed by A, or if you want their amount of collateral. And they need to invest some amount capital I in order to be able to be active on this platform. And this investment project is subject to very standard moral hazards. Again, coming from Armstrong and Tyrol. So they can either pick a highly profitable project with a high probability of success. They can pick a less profitable project, a bad one, which has a lower probability of success, but gives them some private benefit. And they can choose a capital B, a bad project, which gives them an even larger private benefit. And we assume that only the good project could be profitable. Now, the banking system is competitive. So it is composed of banks which aim at providing funding to merchants subject to a breakeven condition. And the banking system has a monitoring technology so that banks can prevent borrowers from choosing the high private benefit, a bad project. But of course, this has a cost, which is denoted gamma p. So in this model, what happens if the platform only prices access to the platform, to merchants and consumers, and only banks provide credit. Now in that case, so if you're familiar with Armstrong and Tyrol, you have the following results that the credit that merchants can obtain depends on how rich they are. And this works as follows. The bank needs to get some cash from the loan in order to break even. But at the same time, the bank needs to leave merchants sufficient pay off so that they have incentives to choose the good project. They need to keep skin in the game. This gives rise to a standard, you know, pledgeable income problem. The merchants cannot give too much to banks. So the banks don't lend too much. So if the merchants have an initial wealth lower than this, we don't see it. Okay. Then this A lower bar, nobody gets funding. And if merchants are rich enough, then the incentive problem is milder and merchants can borrow money without monitoring. Now the difference here is that these thresholds above which merchants can obtain credit with or without monitoring depend on the pricing strategy of the platform. And so if the platform charges higher prices, it will earn more from each merchant, but it will reduce the number of merchants who can access financing. Now on the consumer side, we have the standard idea that the number of consumers present on the platform depends, of course, on the pricing strategy in the sense that only consumers with evaluation higher than the transaction price will enter. Okay. In that setting, very classically, the platform will try to maximize its profit. And here, the only thing that I want you to look at in this objective function is the fact that the financial friction kicks in in this distribution of A. Of course, when you choose a high price, as I said, you reduce the number of merchants which are there. So now let's see what happens when the platform decides to enter the credit market. So the platform can compete. And to compete with banks, it can either decide to offer loans without monitoring or to offer loans with monitoring. And to do so, the platform has a monitoring technology with a certain cost. Gamma P, which can be larger than the cost of the bank. Okay. So what type of contract will that platform offer? An important result is that the platform never has an interest in providing a contract that does not require monitoring. And let me explain you the reason for that. So we start from the situation with only a bank credit, the richer merchants. They are the ones with the payoff in green. They get no credit without monitoring. The less rich merchants, they get credit with monitoring. The poor ones, they get no credit. One thing the platform can do by offering a contract without monitoring is to try and expand the range of merchants who get a credit without monitoring. To do so, the platform can offer loans requiring a slightly lower amount of collateral. But remember, you still have to preserve incentives to choose the good project. So these loans have to be cheaper than the ones offered by the bank. If they are cheaper, it means that all merchants who used to borrow from the bank without monitoring now accept this contract. And so all merchants get a higher payoff. So you see the green surface increases. And it's of no use for the platform. It doesn't increase the total number of merchants. And it costs more. So maybe the platform can try and attract more merchants by requiring an amount of collateral below the threshold of those who are creditors. But in that case, the platform could do exactly the same just by reducing the access price of merchants. So there is no need to use the credit offer to do that because the bank, if the bank wants to have, sorry, if the platform wants to have more merchants, it just decreases the price PM that it charges to these merchants. So the platform's ability to monitor will be key to generate platform credit in this model. Otherwise, credit is a redundant instrument. And let's see now what happens at equilibrium. So again, we start from the situation with pure bank credit. Now the platform will offer credit with monitoring. To expand the merchant base, it will require a lower amount of collateral A to have more merchants accessing the platform. And again, in order to preserve incentives, you need to give those merchants a sufficiently high payoff, which means that the platform will lose money on these contracts. In other words, the platform will need to subsidize these loans to merchants. But if the platform subsidizes these loans, it means that they are cheaper also for those merchants who used to borrow from the bank. So the whole merchants who used to have credit with monitoring from the bank now accept the platform's contract. How far can the platform go? If the platform wants to expand further the number of merchants by reducing the amount of collateral, it needs to set a higher subsidy, up to the point at which it could become even more profitable for those merchants who used to borrow without monitoring to come and accept the platform's contract. And that is never profitable for the platform because then the platform could simply change its pricing. And there's no need to offer credit to capture the whole credit market. So that limits the amount of credit that the platform can offer. And of course, having expanded the range of merchants, now the platform can increase its price in order to increase its profit. So that's the idea of the core result of the paper, which is that if the platform has a monitoring cost that is lower than some threshold, then that platform will enter the credit market. And that threshold below which the platform enters the credit market is higher than the bank's monitoring cost, which means that the platform will enter the market even if it's less efficient than the bank at monitoring firms. And when the platform offers financing, at the same time two things happen. First, it will increase the access fee of merchants to the platform. Because offering credit is a way to implement some form of price discrimination. You want to make your rich merchants pay more. And at the same time, there is some form of financial inclusion because it expands the range of merchants who receive credit. So in this equilibrium, we have first an endogenous segmentation because only the less financially constrained merchants borrow from banks. And endogenously, the more financially constrained merchants borrow from platforms. And that speaks to some empirical results suggesting that platforms tend to provide credit to the more fragile and to the smaller merchants. We can see also that the platform complements the bank credit supply because first the platform targets merchants who were rationed by banks. But it also substitutes banks because it attracts all merchants who could obtain credit with monitoring. And although in the basic model, the type of credit contract we have in mind is one in which the platform asks for collateral and has a lower interest rate than that of the bank, you could also implement the same type of contract by having the price paid by the merchant conditional on whether they accept the platform's credit contract as well. So they would pay a lower price for accepting a credit contract say with similar financial conditions at the bank. That would be an alternative way of implementing this equilibrium. At this stage, I think it's useful to understand what are the necessary assumptions to have platform credits in equilibrium here. First, we need to have moral hazard. If there's no moral hazard, there's no need for platform credits. Why? Because all merchants can enter the platform and the platform charges a monopoly price to these merchants. That is, it charges a price PM such that they make zero profit. The second necessary assumption is the fact that the wealth of these merchants needs to be unobservable. Or in particular, it can be verifiable at the time when you sign the contract, but you can't see it before the merchants sign the contract. If A was observable, in that case you wouldn't need to have platform credit either. Why? Because you could immediately condition the price paid by merchants on the amount of wealth that they have. Now, of course, if they can lie about that wealth, which is the assumption that we make here, if they can pretend to be less rich than they truly are, then you cannot do that. And another point is that, which I mentioned a bit before, could the platform use other instruments like menus? Well, in our case, as long as A, the initial wealth is not observable, menus won't help. Because, again, it's always a rich, you always have to face the constraint that a rich merchant can always pretend to be less rich and accept a contract with a lower platform fee. Now, what else do we do in the paper? I'm not going to go through it, but just if you want to have a look at it. So first we perform some comparative statics to see, of course, if the platform gets more and more efficient at monitoring merchants, it will increase its credit supply. And at the same time, we'll see higher prices charged to access the platform. We perform some welfare analysis. And let me tell you right now that the final impact on welfare of having platforms entering the credit market is ambiguous. Because, of course, it will benefit those merchants who were denied credit because now they can make trades. It benefits as well those merchants who used to borrow from the bank with monitoring because they pay cheaper fees, cheaper interest rates. But of course, it's detrimental to the richer merchants who pay higher prices. Then we consider what happens when we have cross-site network effects. That is, when the decision of consumers to access the platform depends on the number of merchants which are present on the platform. And there we can see that it might benefit even unconstrained merchants. And the last point we look at is the impact of market power, what happens for the credit supply of platforms when their market power decreases. I'm going to stop here. Thank you very much. Thank you very much, Catherine, for the very clear explanation and presentation, but also for adjusting so well to the time slot. So David, now it's your turn. You have 15 minutes for your discussion. Okay. Thanks a lot. So, basically, thank the organizers for allowing me to discuss a paper that I should know about. Okay, because it's building, I see it much more closer to Jonsson-Tirol than to Rochette Viper, as I explained why. But anyway, lovely paper. Honestly, if you have to remind one thing of what I'm saying, lovely paper. Forget about the other 14 minutes and a half, but I'm going to talk. Now, what I think is important for a lot of people is that we want to understand why on earth and what are the incentives leading to big techs and new players in the lending market. I think that's something important for a lot of reasons. And what this paper does is that it tells us that there are incentives. It's not denying that there are others, but it tells us, in my opinion, for a new incentive that I had not seen the debate about. And that incentive is basically what I'm going to call lending for price discrimination. Okay, not for things, not for efficiency, not for regulatory arbitrage, not for price discrimination. And at least I hadn't heard about that argument before. So I think that on its own deserves value. Okay, on top of doing it nicely, et cetera, et cetera. So now, it's a very clear tractable model. It's easy to read. It's beautiful. And basically it has two main players. It has merchants that, as we've seen, half a traditional Jonsson-Tirol model has a problem. So you can take that off the shelf. And you have a platform that basically is a monopolist. So it has price set in power. And it has the option of granting credit. And I think the beauty of the paper is when the monopolist, the platform can actually monitor being worse than banks. The other thing they analyze, I think that's already in other papers. The nice thing is, well, let's put an inefficient platform and still we see that it's going to lend. Okay, I think that's beautiful. And the reason is very simple. It's simple once you understand it. You wouldn't have expected it before. It's because by lending, you actually are able to better price discriminate. Okay, so that's the main reason. Okay, good. Now, of course, I said, I think the paper is timely. I think we have to understand much better why all these big platforms are lending and what impact is going to happen on welfare and on other issues. And I think that this paper disentangles a new mechanism. For other people here, yes, there's no empirics. There's nothing like that. But at least we have to start somewhere. And then other things will come later. So I think that that's, and of course, I'm going to have some comments. So let me just first go very briefly. I think Catherine did an awesome way in presenting. But given that it's home center on, I had to make a leap in trying to say something about it. Okay, so buyers are going to skip. So there's four players, buyers. They are needed. They're taking the Russia beaver. It's a society platform. Yes, you need them. They're asking you there, quote, unquote, boring, not key. That's not where the meat is in this paper, in my humble opinion. So let me skip them. Merchants, big players here. So what do they have? They have a moral hazard problem. And, you know, basically the moral hazard problem is proportional to it. I'm going to explain it a little bit better. But if you want to keep it, you say, well, A is the measure of your moral hazard problem. Okay. And somehow some people can alleviate this moral hazard problem. Doing what monitoring? Which people normally banks? That's the home center all set up. Okay, bullet points two and three would be the home center all set up. What do they add to this picture? They add the platform. And there's one very important thing. Given their assumptions on an observability of A, basically what they're saying is the platform can fix prices, but what's important? It can only fix a unique price for all of the merchants. Okay, this is very important. Okay, because that's what's going to drive this interaction of lending and price discrimination. Okay? And they do it in a beautiful way. It makes sense. Okay, they can only fix one price. Okay. Now, home center all. So home center all basically, in my opinion, is two constraints that basically boils down to one. And that's the incentive compatibility constraint of the merchants. So that's what you have up there. So basically what do we need? Well, we need the merchant to behave, which means that the left-hand side of the equation is higher than the right-hand side of the equation. Okay? Two main elements for that. And this is novel of this paper. One of them is novel. It's the PM, the price that the platform is setting. And another one is the home center all, which is the R. You think about the total loan amount you have to repay to your financier. Okay, something to take into account is that the total amount that you have to repay to your financier is, crucially, that would be the last bullet point there, if the financier monitors. Why? If the financier monitors, there's a cost of monitoring, and then you have to pay the financier that cost, so that he breaks it. Okay, why is this fundamental? Because it's driving you a wedge. Those people that don't need monitoring have to pay less or the things equal than those people that need monitoring. Okay? It's important because this is generating a difference in the profits for those two people. The people that don't monitor pay less in their loans, hence the value of being in the platform is higher. The people that have to be monitored have to pay more on their loans, hence their value is lower. Okay? Yep? Good. Now, what you can see from that equation is basically when is someone more prone to behave, when two things happen. If their wealth is higher, that's the home center all, well, your wealth is high, you have to pay less. And if the price that the platform charges is lower, because the lower the price of the platform, the more you keep to yourself. Okay? Good. Now, this would be the traditional graphical way of doing the home center at least in my classes. You have all the merchants there. If no one can monitor, what do you have? Only low-moral hazard people, those people with high A, are able to borrow. They do it without monitoring, which means that the profits are very high. And this is fundamental for this paper. These people have very high profits. Why? I'm sorry to repeat it again, because they're not having to pay the monitoring costs. Okay? But banks kick in. The banks can monitor. And when the bank monitor, this expands the amount of people that can receive credit. Why? Because it's killing their private benefits. Fine? And you know, there's some people out there that their moral hazard is so high that they cannot receive. But again, the key of the game here is the difference between the intermediate moral hazard and the low moral hazard. Why? Because the intermediate moral hazard received lower profits than the high moral hazard. And that's what a monopolist would want to exploit and cannot exploit because of unique pricing. Yes? As a monopolist, you have two types of consumers. You want to exploit the high-value consumers, but here you cannot do it because you cannot apply discriminative. Yes? That's where credit is going to help you. Now, the platform problem is basically traditional monopolistic problem. Okay? If you increase the price, what happens? You're going to earn more from the merchants you have. But remember, the thresholds are going to move if you charge higher prices. So you're going to have lower amount of merchants. Okay? Now, in this equilibrium, what would happen if you increase prices? Well, basically, as Catherine has shown us, the threshold moves up. The problem is not in the upper threshold. For the platform, the platform doesn't care if someone is monitoring or not. That's a problem if you want of the merchant. What the platform cares is in the lower threshold because he's losing those people, and he wouldn't want to lose them. Well, actually, he can do something not to lose them. Imagine that if you're not being monitored, that's what it also does sometimes called market finance. They call it bank monitoring with a fiber. Let me use market finance. It's cheaper. It's easier for me. Market finance, 2%. Bank finance, 5%. What's the platform going to offer? The platform is going to offer 3% to maintain incentives. Okay? By doing that, what does the platform obtain? Well, basically, it's not very good. Okay? I thought that I could see. But the comparison between the upper and the lower graph. In the upper graph, if I want to exploit the high value guys, I have to lose some of the low value guys by increasing the price. In the lower graph, I'm going to increase the price. So yes, something's going to happen on the threshold up there. I don't care as a platform. Okay? But I'm going to give a subsidy. That's the lower threshold movement. First time decrease the price, I lose some people. I say, wait a minute. Instead of gaining finance at 5%, you can get finance at 3%. So then all these people come back and they actually show that you even expand. It's not that you stay the threshold, you even expand. Why? Because, again, you cannot just pinpoint those individuals, so you just expand a little bit. Okay? Good. So basically, in a nutshell, you have two main effects. One is what happens to the people that had high valuation, that were rich, if you want. Well, this movement is very bad for them. Why? The prices have gone up. End of story. That's bad for them because they don't take advantage of any subsidy of anything. You are extracting those people. That's what the monopoly wants. Who benefits? The intermediate guys. These guys, because the monopoly wants to exploit, the high guys are receiving a subsidy. That's good. Okay? So that's the first main result if you want for me. And the second one is basically what happens because these links to welfare. So basically you have two effects. One is this extensive, what I call an extensive margin effect. By subsidizing the lower threshold, you actually expand credit. Okay? That's what the monopolist wanted. He didn't want to lose these people. So basically, he cannot just pinpoint it. He actually expands it a little bit. In a horizontal T-roll, this is always good. This is just a detail because in horizontal T-roll, if you give credit and the people behave, this actually creates value for society. So that's a positive effect. Is there a negative effect? The negative effect is all these people in between. The people that shift it from being monitored from the bank that does it at a lower cost to the platform that is monitoring at a higher cost. Socially, that's not something that you would want. So that's what puts welfare down. And that's what Catherine said that basically they show that, in essence, welfare could go anyway. They do some parametrications, blah, blah, blah. I'm sorry for the blah, blah, blah. But basically, this is saying, you know, careful is going to be a relation of these two, which was bigger, that was going to drive welfare. Good. Let me skip this. My comments. First comment. I mean who I am, I think this is sort of all. They do something about competition, but it's competition. You want outside competition. Then buyers and the merchants can compete outside. I was generally interested in competition between platforms. Yes? For this mechanism to work, this is price discrimination. It's a monopolist. What would happen in a contestable market? First intuition I got, ah, I'm going to kill the paper. If this is contestable, they cannot price discriminate. There's no way to do it. I'm not so sure. And I cannot prove it. Let me tell you why I'm not so sure. Maybe there's something like a natural monopoly here. Remember, welfare increases in certain circumstances when the platform is able to lend. Well, if the welfare increases, maybe there's a way in which you can reshuffle that welfare increase or not. I'm just saying that I think this is something that might be interesting. And second, if welfare increases, if the equilibrium of the economy would be perfect competition, then you would say, well, there's now a nature to limit competition. Because we want them to have price setting power so that welfare increases. So I think that would be something nice. It's not so far away from what do you do, but anyway, something I like. The second comment, I know you are doing a good job and they explained very well in trying to bridge to literature. And I guess this company is not very fair because I say, why don't you skip one literature altogether? The two-sided platform. The buyers are not doing much, in my opinion, for this idea. Let me give you an explanation. Think about a final good producer, an intermediate good producer, and consumers. I think everything would work the same way. A final good producer has to set prices of the intermediate good producers, the intermediate good producers have more hazard problems, et cetera, et cetera. But David, you still have three. Let me take one out. Think about a house developer that is trying to sell its houses to people that are financially constrained. Then probably the same mechanism applied. I'm not saying throw everything out. I'm saying maybe the main mechanism could even be simplified and then afterwards you say, and platform is a good way to think about this. In essence, they were saying, I understand why you've been on Rochebibers maybe for the simple setup I would take it out. I think this is the only negative comment I have and I'm not sure it's negative. It's up to you. Are there other ways to subsidize? Okay, let me tell you the way I thought about and probably there's a narrow there but I couldn't totally find it. What if what the platform does is, give me your bank contract? So the problem here is in order to subsidize I have to use an inefficient technology. Let me see if I can basically not use this technology. Okay, so what would I do as a platform? I would say give me your bank contract. I observe it to interest rates. If you're paying a five interest rate to the bank which is still monitoring you, I give you a lower platform fee. But it's not that you have to come to me to be monitored. You can go to the bank which is more efficient to be monitored. Give me the contract of the bank and then with that contract I will actually give you a subsidy. Okay, again, there might be, maybe it doesn't work because there's ways to overcome the contract in some intelligent way but two seconds. But if not, I mean the other reason, well, the partnering between banks and fintech but that's not the real problem. The real problem is could I create a contract which is based on bank observable interest rates? Okay, anyway, really super interesting paper and very thought provoking. Thank you very much. Thank you very much, David, also for very efficient in your delivery. So let's, we have some time for taking questions from the audience. Yes, please, here. I would ask you to stand up and to identify yourselves, please. Yeah, Jean-David Sego from the ECB. I really like the idea that the big tech are entering the landing market for price discrimination. Have you thought about extending it to a dynamic setting? Because in such setting, in principle, nothing would prevent a big tech to get the information about which merchant is rich and which merchant is poor and then altogether withdraw from the landing market because they have the information to do the price discrimination. They don't care about the landing market anymore. There's probably a quick fix in a dynamic setting but I just wanted to have your view on this. Thank you. Yes, please go ahead, Catherine. So you're right that the, you know, the main reason why you want to price discriminate this way is because you can't observe who's rich and who's not rich and people can always pretend that they are poor than they actually are. Which is, you know, to some extent, valid maybe for these merchants but now if you can secure, if you can be sure of their wealth then you have, as David said, you can still price discriminate but in other ways. And so then you might provide less credit and provide more, you know, pricing based on this information that you have. So you would have the same welfare effect if you want but who are different instruments for sure. That's, but you would still have this form of price discrimination but less credit. That's totally true. You could have also other dynamic effects if you're thinking, even in this setting, keeping, you know, the assumption about the fact that the wealth is not observable. You know, in the model, FinTech, sorry, big tech credit, substitutes bank credit for the monitoring part. So you could imagine a dynamic model in which platforms get more and more efficient at monitoring. And by doing so, they could provide more and more credit to merchants because now they have both an interest in price discriminating but they would also have an edge in terms of efficiency and that would kick the banks out of that market more quickly which could have, you know, even some effects at a more systemic level. You know, if that type of monitored loans are now more and more in the hands of platforms and less and less in the hands of financial institutions then that could create some issues regarding, you know, regulation and systemic risk. So with a dynamic model you could capture that as well. Thanks for the great presentation and fantastic paper. A question that relates to David's last comment. Have you thought about the market for data? Because I guess in your setting the platform would be interested in purchasing, let's say credit scores from the bank to identify who's a good borrower who's not to use that for price discrimination and that could have an interesting effect on the behavior of the bank as well as the platform and some new interesting avenues to explore. Thanks. You're absolutely right and David was right. You could, so some of the inefficiencies that arise in the model in terms of welfare are due to the fact that you have this platform providing credit which is less efficient so you increase your overall monitoring costs. Now, if you could thanks to data thanks to contracts with the bank link a bank contract with a lower price which is what you suggest you could implement exactly the same you know the same result that you can increase prices and expand the merchant base. You would still have an ambiguous effect on welfare though not because you're less efficient at monitoring simply because you have more monitored credit overall compared to the case with banks because some merchants who used to borrow from the market as you said they will now accept the bank contract just because they have a lower price on the platform so you would reduce it's true that the inefficiency of having this less efficient platform monitoring will not be there but you will still have too much bank credit compared to the situation the platform does not contract with the bank so there could still be a welfare loss Okay, very good so I think we can now we should move now to the second paper of the session thank you very much Catherine and David Thank you So the second paper of this session is going to be presented by Marco Di Maggio from Harvard Business School the title of the paper is Invisible Primes FinTech Lending with Alternative Data is a paper written together with Dimitru Randa-Wikara from Louisiana State University and with Don Carmichael from AppStars Operations and it's going to be discussed by Tobias Berg from Guetta University from Ford so Marco please the floor is yours Perfect so thanks a lot for having me pretty excited to present this paper together with Dimitru so just to show you the connection with the previous paper I'm going to show you empirically what some of the FinTech lenders do in the market the way in which they operate the way in which they compete with traditional intermediaries and the way in which they potentially use alternative data so the overall framework of the paper is just thinking about how credit markets have been changing in the last few years in particular the introduction of FinTech lenders as disrupted credit markets to the point that at least in the US they have now placed an important role in credit markets in fact if we think about the one that has the largest share in mortgage origination is quick and loans which is under the umbrella of FinTech and so thinking about why now they have a leg up on the banks and what type of technology they are using is sort of the focus of the paper in particular we're going to focus on the effects to households and some of the common wisdom around the use of and the emergence of these new technologies being about improving credit taxes up to now only the borrowers that have high credit scores are the ones that are able to rip off some of the benefits of having access to for instance multiple credit options and here we are thinking about instead the left tail of the wealth distribution what we are going to call from now on the invisible primes the ones that potentially are still credit worthy but they are somewhat cut out from the credit markets and we're thinking about how to what extent FinTech firms might potentially help them so how big of a market is there for this invisible prime the number that I like to sort of quote here is that think about the people that have defaulted in the US well four out of five Americans have never defaulted on any type of loans conditionally on having a loan but just less than half have actually access to prime credit the ones that lower cost so it seems there is a discrepancy between the real credit worthiness of these individuals and their ability to have access to the best possible options in terms of credit products and this is would be improving credit access would be good potentially for both borrowers and lenders lenders could lend twice as much having lower default rates if we think about how lending happens especially in the US is all based on few statistics the FICO score, the credit score is the main one just to give an example how this has been used in the US the GSEs have a minimum credit score of 620 so you are credit rationed if you have a lower credit score and both in some of the new emergent lenders Marcus which is part of Goldman Sachs and Santras which is a traditional bank have a minimum credit score of 660 for unsecured credit loan for instance and you know maybe they are right to do this because if you run a regression looking at credit scores on the full probabilities of both credit cards as well as mortgages which we have done here you see that higher credit score corresponds to a lower default probability the issue is, is this the whole story or are we missing part of the information we need to actually assess the worthiness of these individuals so these are the how big are the credit in the invisible in the US this is just by age category these tend to be younger folks and this might be invisible because either they don't have a credit score at all or they have a stale credit score because they didn't really use the credit market lately or because they have a thin credit file meaning they only have a small credit history but they haven't gone enough information to build a credit score so these tend to be all over the wealth over the age distribution but definitely the younger are the most important one then if you look in terms of income and minorities are the ones that you would expect these tend to be low income individuals tend to be black and Hispanic and so there is also a question of well maybe there is an inability of the FICO score as tradition has been fought in predicting credit worthiness for some categories in the population and you know this has been sort of taking over by policy makers, the CFPB have said maybe we should be using alternative data because we want to improve access exactly for these folks this is a quote from Richard Kodra he was the director of the CFPB and so these are some of the questions that we ask how does using alternative data and better models affect credit access what type of data are actually helpful here and are consumers ultimately better off and who wins and who loses potentially using a different credit underwriting model and finally do alternative data actually reduce credit access for some individuals on the wealth spectrum so what would be an ideal setting to answer this question first of all you would need the underwriting model to use alternative data I think the hardest part is actually identifying the counterfactual so you will need to observe in the data somebody who applies to different platform ideally to a traditional lender and a fintech lender and then really know that the extent to which they've been accepted by the fintech and not the bank is due to the underwriting model we cannot really observe that usually in the data and so whether the same borrowers would have been accepted or rejected by an alternative lender is sort of the big counterfactual we need to estimate and then you will need a lot of other things you will need to observe the landing decisions make the difference between the ones that get funded and the ones that are not funded as loan applications you will need to understand where the underwriting differences come from so you will need to understand the underlying model and then you will need a panel data for both the borrowers that have been funded which potentially is doable but also for the ones that are not funded which is much harder to observe because they didn't get the loan so it's harder to actually observe later on what happens to them we are setting answer most of these ideal characteristics we focus on a company Upstart which is one of the largest fintech in the US why Upstart is helpful first of all checks the list in terms of it's a lender that uses alternative data I will be talking about this in the next 20 minutes and he uses also an alternative credit underwriting model and then this is attracted the interest of policy makers in the US the CFPB was very worried about the usage of alternative data and they had to issue what's called the no action letter basically telling them you can go ahead we are not gonna enforce you what is the main characteristic of the data that helps us identifying what we are looking for is we observe the alternative model and this alternative model has been developed by the CFPB exactly to understand whether this Upstart, this platform will be discriminating against borrowers they built a new model which is sort of the an arms race between the CFPB and the model provided by the platform to understand what are the borrowers that are excluded what are the ones that are included when you change the information set of these models and so it's something that we don't have to come up with but we actually borrow from the regulator right away and then the second thing that is important for us is also that we observe a panel for both the ones that are funded and the ones that are not funded as loans and this allows us to do a number of things in the paper what are the main results I'm gonna be going quickly towards through many of the results so I'm gonna make sure that at least some of the main takeaways are clear the paper is sort of structured in three ways first we explore what is the model what is really the what are really the main drivers of this model then we go through what are the differences between the model that is provided by this platform and the ones that is instead provided by the regulator and by another large bank in the US so we are really looking at does that lead to a credit access does that lead to lower rates for borrowers and then finally we go to what are the effects on the households can we estimate whether the households getting this credit are actually better off so what do we find we find the platform model does not really rely on the FICO our performance traditional methods and there is a composition of both the model as well as the data so both are important two thirds of the way in terms of the improvement comes from the data one third comes from a better model then we can say that the main data alternative data that are helpful it's shocking that the banks don't use this this is not rocket science but it's education, employment, digital footprint these are things that are incredibly predictive of credit worthiness and are not discriminatory at least in the US what we find, how big of an effect we find is between 15 and 30 percent of applicants with low credit scores that are accepted by the platform are funded by the platform would have been rejected by all the benchmarks from the regulator as well as from the banks so there is a huge effect both on the extensive margin as well on the intensive margin so conditional being accepted by both the platform as well as by traditional banks we find that being accepted by the platform led to about $8,000 in savings over a $10,000 loan over the life of the loan so there are huge effects both in terms of the intensive as well as on the extensive margin and finally we do find that the borrowers that do get this additional credit are much more likely to purchase a home later in life much more or less likely to default on any other type of credit card or any other type of loan product and they are more likely to experience an increase in the FICO score as well so they are better off among some of these dimensions that I just explained the related literature just a shout out to my discussant has been one of the first papers looking at digital footprint in this area but in the interest of time I'm going to skip this so let me go directly to just the data we cover 2014 2021 with about 700,000 loans that are funded think about these borrowers over the seven years, 2.7 million that are non funded what we can observe for these two data sets we can observe the credit score and the credit report at origination for both and we can observe the performance of the loans for the ones that are funded obviously we cannot observe the performance of the loans for the ones that are not funded and then we observe everything else for both of these two subsets we observe the outcome of the model we observe how the model has been run we observe the model from the regulator we observe what are the inputs of this model in particular the alternative variables this platform has been growing a lot the data stops in 2021 the first graph shows you the number of loans that picked to about 300,000 loans in 2020 we get to half of 2021 and they were sort of on track to do exactly the same so one thing that I want to mention is that we also have the funnel how these get to the platform so if you're worried about selection why should I be applying to upstart rather than applying to a different bank potentially might drive some of these effects the nice thing here is that we observe most of the borrowers that get to this platform by applying to what's called credit karma which is a loan aggregator so they apply to all the lenders at the same time just that the offer that they receive are the ones from this platform because they are the only ones that are actually offering them any product at all this market has been growing a ton together with upstart so it basically more than a triple in the last 10 years between 2010 and 2020 to 150 billion dollars in the US so just the main difference between the funded and unfunded the funded tends to be highly educated slightly higher annual income although they still have a lot of liabilities on their balance sheet and they have a credit balance they already have access to loans and less likely to be early employees now let's compare let's do some fun about the platform model and the three benchmark as I said our main benchmark because it's the one that is the most conservative is the one provided by the CFPB this is a logistic regression based on the same variables as the platform model except for the three main one that I mentioned the education, the employment history and the digital footprint it's complex compared to what we got from a traditional lender so we also got the model from a traditional lender which is between me and you it's anonymous, it's a two by two matrix it's only based on credit score, DTI and loan amounts and this is one of the top 25 banks in the US the platform model instead of the 1600 variables to predict the probability of default includes the traditional credit reports variables and a mixture of those interactions of those plus the alternative variables and so as you can see there is a difference both in terms of the model as well as the data we're gonna then decompose these two and then with the third benchmark just using the credit score now, first of all are they any different? Yes, there is a huge difference what it shows here is the traditional model output on the x-axis and the y-axis you have the probability of default predicted by the platform what those distributions show you is that even conditional on the same probability this is from the traditional model there is still a huge variation in the probability of default given by the platform what that means is that the platform is giving you more granularity than what the traditional model is giving you and in particular some of the even high credit score individuals even above 700 end up being rejected by the platform because they have predicted a probability of default above 50% if you compare this with the large bank model both for denied and approved as you can see the distribution is even wider and if you compare this with credit score sort of the same thing what you see is that the shift towards the higher lower probability of default for higher credit scores that's natural so the two things are correlated but still even for people that have a credit score of 740 you still have probability of default that are above 50% so they would be denied if you look at the performance of these models there are a couple of different ways one can use one is why don't we look at the faults of these loans as predicted by these different models and so traditional model output again X axis, Y axis is the probability of default and these are actual defaults so what you see is that the darker areas are borrowers that have defaulted on their loans so even the ones the traditional model will say are very credit worthy and up saying that instead would be defaulting given the probability of default of the platform same for the credit score is even more concentrated the best difference between the best way of really capturing the difference between underwriting models is really looking at the area under the curve this is 50% if it is a toss coin the higher they are under the curve the more you can differentiate between the ones that will default and the ones that will not default and so what we plotted here are these areas under the curve for the three different models just the credit score, the traditional model and the platform model as you can see there is a big difference the credit score is really a coin toss as an area under the score of 51 meaning it cannot really differentiate between the ones that will default and will not default this is based on the sample and then the platform instead has 20% higher area under the score both for low income people low credit score people as well as high credit score people then the final question about the platform model is what are really the key features of the model sure they can tell us that this is all about alternative model but can we actually test whether this is true or not we do one thing informally which is the recursive feature elimination which basically is going through a random forest looking for the variables that predict the probability of default the most and basically there are 15 variables that predict the probability of default the most in this model among the 1600 I told you are really the ones that make most of the variation here and these are the level of education type of job loan purpose and some of the digital footprint let me skip the table the other interesting thing is can we actually put these results into a different context thinking about okay is there an exchange ratio between education and income between education and FICO score one nice way of thinking about this is that well if really I'm putting a lot of weight into education then it should matter a lot whether you have an advanced degree or high school okay it should matter a lot in terms of how should I compensate for it one exercise we can do given the model is that to go from conditional on credit score loan amount and age so you look exactly the same except if you go from high school to an advanced degree you have to compensate that you need to earn about $107,000 more in terms of probability of default as safe as they were or conditional on age income amount to go from high school to advanced degree you will need to have about 40 points higher FICO score as well okay so this is another way of thinking about how important how a big deal is opening this Pandora box of using alternative data now can we decompose between data and the model we did this with area under the curve and what I I can tell you is that we re-plotted the same exactly the same thing for two different models the traditional model now has been argumented with also the alternative data and so now the only difference between the traditional model and the model provided by the platform is the underlying model itself so the logistic regression of the policy maker against the machine learning algorithm used by the platform okay the other one is doing the opposite of looking only at the extent of the alternative data comparing the usage of alternative data within the same model of the platform and what I told you is that about two-thirds of it is actually coming from the data one-third is coming from the model so you don't need to actually big message for both policy makers as well as practitioners you don't need to restructure your whole IT department if you believe these results you will only need to actually include alternative data which are again not rocket science to improve your underwriting model okay let me go into one thing that you might think about which is well there is an issue about the data that you are looking at these are data that have been either funded or disqualified by upstart maybe these alternative models, the alternative data don't really predict the probability of default outside of this sample so what we did was let's look at everybody who has such a low credit score that also this platform would deny credit too so these have not gotten any credit by anybody can we actually look at whether this alternative data predicts their probability of defaults yes they do so having college degree, advanced degree or using for instance a particular computer so even some of the digital footprint predict the probability of them being defaulting on any type of delinquency or credit cards 90 plus delinquency which is the main measure of delinquency in the US for personal loans now the main results about the extent to which having a better model and more detailed data helps is let's look at what would have happened in the counterfactual world and this is what we do here we basically are plotting we are regressing the approved the approved loans against the large bank model, the platform model the traditional model and what you see here is that basically the traditional model between 620 and 660 would have rejected the probability 1 this is conditional instead having been accepted by the platform the traditional model by the large bank is a step function because as I sort of hinted to is really a 2 by 2 matrix between DTI and credit score so it would have been rejected again up to 690 all the applicants and then we'll jump up and still being below the platform model what that means is that difference between the ones that would have been accepted by the platform and the ones accepted by these traditional models are all about the extensive margin these people are getting credit that they will not get otherwise from other lenders if we do the platform approval rates for the ones that are rejected and very similar effects meaning these are even among the ones that are rejected by the platform you see this difference between the models now let me go directly to the intensive margin so we can do the same thing which is this was one of the main questions Xanti for us was ok how can they be competitive but maybe they are giving credit and they are just losing a ton of money maybe they are giving credit but they are charging such a high rate that the banks don't feel to charge to these to these households and so we looked at the interest rate the APR for these different borrowers and what we plotted here is the platform model the lighter blue and the traditional model is darker blue against the FICO score and basically what tells you is that normalized at 800 what that tells you is basically the difference the spread that traditional model will charge to somebody conditional giving him a loan with respect to what the platform will charge and so irrespective of the credit score they will be saving a lot of money by going to the platform rather than the bank now in terms of who benefits the most apart from as I said the ones that are really loaded on education having an advanced degree matters a lot I wish my bank would give me the mortgage based on the education it doesn't income doesn't really matter as much but it matters the employment history so really the one thing to think about is that the stereotypical borrower of a platform like this is for instance an immigrant somebody who has a very high education but doesn't really have a big credit file he's just building credit in the US the banks is not willing to lend to them but they are very good credit they have a nice job, nice education they are gonna be able to repay it let me go to the last piece which is does credit access improve outcomes for these borrowers this is the ultimate question and in the end what we the way in which we do this is by using the facts that there are some discontinuity in the data some borrowers are disqualified automatically given the DTI so that's the income ratio for them and so we can use a regression discontinuity framework to see what are the effects one year down the road, two years down the road for these individuals was it actually good for them to get this credit and let me tell you just the punchline the punchline is that we see them defaulting less on everything else and more importantly we see them purchasing a home as well so it seems like they reach the stage where they can actually go out and go to a traditional lenders so this seems like the transition from being invisible to become visible for everybody else also who are the ones that lose the most from all of these again let me just give you the punchline the ones that are for instance retires are high income in the low income individuals but high FICO score individuals these tend to be the ones that are not approved by the platform but approved by the traditional lenders so you can think about this as the FICO is really a backward looking measure so it's looking at the fact that you didn't default so far but it's not incorporating the fact that you are not earning much income now while the platform is making that difference as well and between being actively in the labor market and the ones not being active as well let me go to the conclusion I think the conclusion is all about what the police maker can do given these results I think can do a lot of things because in the US the fact that everybody relies on the FICO score is induced by the regulator is induced by the regulator because those are the requirements by the regulator are all based on cutoff of the FICO score this is at least these results and the literature starts showing that it's going to miss a lot of opportunities in terms of improving credit taxes in the US thank you thank you very much Marco so thanks a lot first of all for inviting me to discuss this really interesting and current paper actually I'll start with a very short summary the paper is about the effect of credit scoring with alternative data on credit access and on real effects setting upstart a major fintech based lender in the US with data coming in particular from 2019-20 in the beginning of 2021 and the key result A on the model part using alternative data improves credit scoring number two on the credit access side you show at least you claim in the abstract so far that the traditional model 70% higher probability of rejecting boroughs and the key beneficiaries are these invisible primes with a low score and a short credit history and then you also look at real effects and you show that there are better economic outcomes for the affected boroughs so overall I very much like the setting because this is about a major fintech lender in a developed economy and if we look at the fintech just so far there are a lot of setups which look at very very specialized niches and here we really have a major lender in a developed economy and like with a really big and nice credit portfolio you can look at. The other thing that I think is really beautiful is that you have results along three dimensions and many prior papers actually focus on one or two of these at most. So you both look at the effects on credit scoring does alternative data help to improve credit scoring? You can also look at whether this improves credit access because better credit scoring does not necessarily one-to-one translate into better credit access and you look at real effects for the affected boroughs. The paper also has a quite catchy claim that I've also seen cited also frequently in actually other papers or policy papers. It's about invisible primes and it's about a 70% higher credit access that's kind of the benchmark number you have in your abstract and I'm always trying to be very direct in my discussions and so I have a clear statement up front. It's a very nice paper but based on the paper that I've seen from you it's somehow significantly short of the quality that I expect from a difficult marketing much of paper. I'll say that bluntly directly up front and the interesting part actually is because it's a work in progress paper that at least based on two of my comments that I will do your presentation was already different in a way that I liked very, very much actually. So I think it can turn into a very nice paper but with some of the claims that you currently make in the paper that I've seen they might be a little bit over exaggerated. Let me give you some background first of all on the prior research. Marketplace lenders actually did have a rather poor performance over the last 10 years. There are a lot of old papers on Prosper and Lending Club but they have not performed very well these marketplace lenders. The effect of FinTech lending on credit access has been limited in developed markets so the prior papers actually show that effects are not so big when you look at the effect of these FinTech lenders in developed markets. And you have your own paper on the US FinTech lending where you actually show that there's sometimes a short lift reduction in the cost of credit but there are actually negative effects on default rates in another FinTech lending setup in the United States. And so this paper claims 70% increase in credit access and better real outcomes. So if this is true I think this would be a significant shift in most people's priors actually. And this is also I think the paper is potentially fascinating and interesting even though I have a couple of quibbles as I'll say in the following. So result number one. Your result number one is on credit scoring and the fact that alternative data can improve credit scoring. There's a host of papers on this in developed market including the US already. IA et al for example I've looked at Prosper data, you can look at the Lending Club data, I've looked at just the digital footprint which is a sub part of your alternative data. The paper by Agraval et al uses at mobile data and all have effects ranging at let's say roughly plus 10 percentage points in the area under curve in the discriminatory power. And your study finds I think in the baseline setting that this model by upstart improves the discriminatory power by 11 percentage points relative to a credit score, a credit bureau score. And I find this highly plausible, it's also backed up by prior research. I think the interesting part is really the variables that you show employment history and education that they matter so much. I think this is the most interesting part in this first part. Your traditional model I think seems to perform somewhat worse than indicated in other studies. Your traditional model only improves by 6 percentage points relative to the credit bureau score. I would at least be interested due to the sample, the data, the methods from other papers I would know more things like maybe a plus 10 percentage point improvement for these type of traditional models. Now therefore I think result number one, I find highly believable. I find it interesting to see which variables are driving it and it's also in line with the literature so far. Result number two is the improvement in credit access. In the version of the paper that I have received, the abstract very prominently said 70% increase in credit access. Roughly speaking. When I look deeper into the paper, it actually says okay for borrowers with a credit score between 600 and 640 who are granted a loan by upstart, they would have a 70% probability of being rejected by traditional lenders. And that's obviously a big difference because 600 to 640 is 5 to 10% of the US population. The acceptance rate by upstart in this segment is roughly 20% and at least based on a naive web search there seem to be some lenders who are also willing to lend or give credit cards to borrowers in the 600 to 640 area. So I think there's an easy fix for it. Make it a more balanced and maybe more economically relevant claim and what you have highlighted actually in your presentation is this I think 15% number which I find kind of I think much more economically relevant and so therefore this comment is more or less already dealt with then. I make this comment because I see many papers actually or many people probably reading abstracts and then using exactly these claims the 70% numbers and I think that we should look at it. So thanks for kind of addressing this comment even before I made it. Now result number three on the positive really facts. You write on your own paper in this paper here that this is contrary to the results of your prior paper but actually without really explaining. So I'm really interested why is it different. You find an improvement in default rates, an improvement and a higher likelihood of purchasing a first home and in the major part of the paper you something which is similar to a matching strategy or controlling on for a lot of factors. And what you write in very much much economist speak is one concern is that borrowers are likely to differ on unobservable characteristics that might be correlated with both the probability of being funded and later outcomes. And I try to translate this kind of in simple plain English. Yes, if you attend Harvard then hopefully your life will be different not only because you get an upstart loan. Hopefully Harvard does something more to you than just offering you a better ability to get an upstart loan. It will do hopefully something else to improve outcomes in later life including default probability first homes and the like. Now interestingly again in this presentation you have highlighted this matching strategy not a lot which is I think a good idea because it has really some significant shortcomings. What you highlighted is this regression discontinuity which is a quite interesting setting here actually. So you say that upstart also uses very simple cutoffs in this case a 50% debt to income cutoff. If you have a higher debt to income then 50% you are not funded otherwise you are funded. You then use which I think is clever in RDD to kind of roughly compare those with the 49% debt to income to those with the 51% debt to income to get kind of the causal effects of an upstart loan on later outcomes. And I like this and I think it should be expanded. Now one problem or another one I have with this is that when you look at this sample per se these are not invisible primes anymore. These are people with the 50% debt to income ratio so these are highly indebted borrowers. When I looked at your statistics in this particular sub-sample at the cutoff the average borrower has $80,000 in debt on outstanding so they have been served by the traditional financial industry. And they have a lower credit history and a 670 score on average. So while you get at a very nice causal effect here it's no longer about invisible primes here. I think what you can do actually is you can do an analysis exploiting the heterogeneity here to some extent. Yes they will all have a high debt because they have a high debt to income ratio but at least you could probably look at a sub-sample of these who maybe have a short credit history but still have already maybe a higher debt ratio outstanding. I think that would be interesting because otherwise there's a little bit of disconnect between this invisible prime claim and then this causal effect here which is more based on borrowers who are already highly indebted. There's one elephant in the room and in some sense you might be a little bit unlucky here because things have unfolded in the last one and a half years a little bit detrimental to upstart. So the key variables that you find that matter are education and employment history. And so I was wondering why is that not used by traditional lenders? Why don't they use education and employment history? And it might either be that upstart has a novel methodology to validate this because everyone can claim they've been at Harvard I think it would be interesting to provide evidence if this is the case if they have the technology to finally validate these parameters with a lower cost than banks could do. Number two, you alluded to this a little bit, the US is a complicated regulatory environment where many lenders are very wary of using data that might violate fair lending acts. And the third one might simply be due to the fact that it's just not robust to the economic cycle. And when you look at upstart's market capitalization or share price since your first version of the paper at the end of the data sample it has dropped by 90%. It went down from 30 billion market cap like really upstart this will be the next big lender that will take over the US banking system to less than 3 billion now. And I was looking a little bit at the reasons why this actually happens. So first of all I found many articles that claim that upstarts violate the fair lending act. And there was even initiative by I think many prominent senators including Warren for example Brown and Harris and the like who said this needs to be looked in detail because you discriminate against historically black colleges. And it's not only an allegation but there was a report that was written and then upstart said ok we'll improve our models I don't know whether improve means that they get a better discriminatory power or actually worse discriminatory power because you know they don't want to take these variables into account directly. Number two the economic cycle. I just looked at the Wall Street Journal as an example you could look at other journals probably as well or other newspapers and so it seems that a little bit upstart was hammered over the past one and a half years. And if you look at any reports of upstart you find similar information actually. So in their 2022 results they say revenues has decreased by 52% and lending has gone down by 62%. If you look at the latest available in June 2023 there's again a decrease of 40% in revenues and a decrease by 64% in volumes and they moved from profit-making to loss-making actually to a significant degree roughly from plus 60 million to minus 60 million. And I've already talked about the market capitalization which went up down from 30 billion to 3 billion. So where does this leave us? I think it's a highly relevant topic and I very much like that you focus on major market and not these kind of niche papers that others including me have written with kind of a niche setting but it's really a major market here and that it's really broad in scope. You look at scoring credit access and actually real outcomes. The 70% claim we talked about it I think you've already actually revised it and I think the elephant in the room is really if this is performing so well why are these problems so big in 2022 and 2023? So if you have additional data from 2002 or even 2023 I think it would be really interesting to see what went differently after these fair lending adjustments and how did upstart actually cope with economic cycle? Actually why did they cope rather poorly presumably with this economic cycle? Thank you. Thank you very much Tobias just in time so I will open the floor for questions but Marco you may want in the meantime to address if you wish some of the issues raised by Tobias. Yeah, we'd love to take 30 seconds before the question so thank you so much for the discussion. This was great so in terms of the 70% number that was a relative number with respect and maybe we should definitely be clearer on the abstract if you think about what the picture that I showed you the traditional model basically staying at zero in terms of approval rates between 600 and 680 while the platform model will give you lending the 15-30% is sort of the average across the whole distribution. The other thing that I want to make sure of is think about why the RD which is focused on this 50% predicts that people have more than $80,000 loans it seems a lot but think about the stereotypical person that I told you is these are all student loans my students have more than $80,000 in student debt and so the reason why they are still invisible is because they are stuck with the student loan they cannot access to anything else because they are highly indebted but they are indebted not because they have 10 different credit cards but it is because they are actually financing their education so that's very common among these folks why the banks don't use this this is a big puzzle for me as well well the CFPB until 2022 actually granted the no-action letter so they show that it's not discriminatory so they could do it the banks could do it and can they do it technologically yes actually education and prevent history gets verified in the US by Equifax in the same way in which you can get your FICO score it's pretty scary but the credit bureaus in the US collect much more information what we can observe and we know of they why because they sort of you know if you go to get your credit report you always say who is your employer your income because it works as a registrant for 5,000 large employers in the US about the upstart market crash hopefully there is no casual causal inference from my paper to market cap of upstart but I will say sort of it's relative to the fact that their funding system is all based on having investors that buy these loans as interest rate spikes up all the fintech markets sort of collapse I will say at most questions Simone please European Central Bank I would like to follow up on the question on the fair lending and you said that you're not using these variables which are considered unethical do you have these variables I guess you know like sex or religions or race and you know if you put them in how would your results that the borrowing history or the education are still the main explanatory variables for your results they are so scared of even touching any variable that will put them under the cross size of the regulator they don't save any of those the only stage in which they had to save that info collect and save that information was to show to the regulator and the regulator was running the model itself right so it's not like the upstart had to the platform had to show this but the regulator was doing this or race and showing whether the probability of default would discriminate across any of those dimensions but it's not in our data unfortunately or fortunately to stay safe Luke over there I think we're all hunting the same elephant here the fair landing because I start from like this I mean you find this huge effect so either banks are lazy stupid or plain scared and I think it's the letter because of this equal opportunity act which stipulates that cannot look at race and religion and sex and I think if you just take the first variable you don't need your data just look at national data you'll find that it perfectly explains variation in education across the US and so well what this lender is doing is not explicit discrimination they can find a lawyer that will protect them it is implicit discrimination that banks are probably very scared to do and indeed over the history US banks credit scoring models have deteriorated because they have been taking out all these variables either the explicit or the implicit ones and they laugh at the FICO score plus one maybe and so I do think that even if you don't have it in the data set just pull out the more general national statistics and see if there's a strong correlation or not I think this question will come up so I think the regulators for granted so if this CFPB says the random model says it's not discriminatory it means that conditional on the other 1500 variables does not predict different appropriate to the faults between race, sex and religion you can say sure having the Harvard degree not having the other degree is discriminatory but that's an unconditional statement the model gives you the conditional statement on everything else and that's what you should check for discrimination or not I think so we know that loan officers have some discretion how do we know that this discretion is not based on this missing variables like education sure and it definitely might be and within the realm of paper that will sort of bring our estimates to be lower bound because then you are a competition between the model plus some soft information which we can really capture in some sense if you think about the extent to which the difference are there obviously the bank can catch up with some of the soft information but we cannot really really look at this one thing let me go back to look because the other thing I just reminded myself that I haven't shown what it's in the slides is that we do one thing regionally so we look at zip codes because we can observe the borrowers level but we can match this to zip code information so we look at for instance the fraction of people living in a certain zip code that are minorities and what we find is that those are actually more likely to be winners rather than losers so those are places where you would see an extension of credit more preponderant from the platform than from the traditional traditional vendors. It's not perfect because you would like to have the borrower level but at least it tells you that there is across zip code you see that it helps the left tail of distribution more. Very good any other question, comments? If not we can conclude here by thanking again Marco and Tobias for excellent presentation on this category.