 Okay, so we have further two papers in this session, in this day actually. So it's 25 minutes for presenters and 10 minutes for discussions. So we start with Valerio, and the paper is called Risto Buffer, setting a signal in structural capital buffers through bag stress test. So it seems super relevant for the discussion we just said in the first part. So thanks a lot to move the slide. So thanks a lot for giving us the possibility to present this paper. This is a joint paper with Cyril Quaglié, that is also at the European Central Bank. What we do in this paper is to propose an approach that we call Risto Buffer, to jointly set signal and structural buffer through bank stress test. Risto 3 introduced the distinction between two types of buffers. So we have cyclical buffers on one hand and on the other hand we have structural buffers. So the cyclical buffers are the buffers that vary with the financial cycle. So let's say within thatness and other factors of the financial cycle. One example of that is, for example, the counter-cyclical buffer. Other buffers, instead, they remain constant across the financial cycle, so the structural buffer, and they cover different types of bank's vulnerability. Importantly, they are also set at a lower frequency. So we can think, for example, of the capital conservation buffer, the GC buffer, or the P2G buffer. And also, for example, one example could be also the non-cyclical component of the counter-cyclical buffer. A key important caveat here is that the structural buffers, like, for example, the capital conservation buffer and the P2G, can ensure resilience against economic downturns, but are not quarterly set to counter financial risks. So for example, if financial cycle conditions evolve, these are not going to be the buffer that are going to move, but we are going to also only see a movement in the calibration of the counter-cycle buffer. In prudential policy, buffers are often calibrated with stress test models. So we have macroeconomic models that are generated and used in bank stress test models to obtain capital shortfalls. Based on these capital shortfalls, we are going to calibrate the capital buffers. So this raises a couple of questions. So first of all, starting from this shortfall, how can we disentangle cyclical and structural buffers? So which part of the shortfall is going to be devoted to the cyclical component of these buffers and which part is going to be instead devoted to the structural? And also, another point is, when we're running this stress test, how does the evolution of the financial cycle can affect the calibration of the cyclical buffer? So we're going to see variation in the cyclical component and in the cyclical risks. How are we going to map this into our stress test framework? And also, the clarification about this issue is not just philosophical because it also matters, for example, to avoid the risk overlap between cyclical and structural buffers. So this, for example, could bring to an inefficient and not transparent use of the capital framework. What we do in this project is to provide a conceptual framework to think and to calibrate cyclical and structural buffers. We call this risk to buffer. So the idea is that we're going to map each risk to a buffer, each risk level to a buffer. And it is based on the integration of a nonlinear macroeconomic model, the cyclical amplifier. We call it like this and we're going to see how and why. And we are going to integrate it to a bank stress test model. So in the first step, we generate state-dependent scenarios. So the macro model provides these scenarios whose severity depend on the level of cyclical risks. And the idea is that the higher the risk, the higher the severity of the scenario. And multiple scenarios are generated by using the same set of shocks, but for different risk levels. So I have a set of shocks, let's say housing and credit shocks that hit the economy. And the propagation of the shock is going to vary with respect to the risk level. And the higher the risk overall, the higher is going to be the amplification of the shock in the economy. So producing a more severe scenario in case of higher risk. Then what we have to do is a policy decision, which is we need to fix a reference level. So we have to fix a level of risk that we consider as the reference risk, a sort of medium run, a sort of reference risk for which we can say the losses that we are going to obtain under this risk are going to be covered by the structural buffer. And the additional losses that we are going to generate by the amplification of the cyclical risks in the model, these are going to be associated to the cyclical buffer. Let me just show this with an illustration. So we have on the left-hand side the evolution of the capital losses in our stress test model. So we think that we had our macroeconomic scenario with different risk level, different severity. So again, the higher the risk, the higher the severity of the scenario, obtain always using the same set of shocks. It is going to be our nonlinear model that is going to deliver this different severity. Then we use this severity in a stress test model. And we obtain different capital losses. So the capital losses, for example, under the minimum risk, are going to be like the blue line, so smaller. And then the higher risk, for example, the median risk is going to present higher losses. And let us assume that the current risk is higher than the median risk. Then we are going to have a larger losses under the current risk scenario. So if, for example, we fix here, there is two buffer for working this way. So we fix a reference level with respect to an historical risk. So we say, let's say that the median risk is our reference level. So the losses, like the yellow losses that we are going to obtain under the median risk are going to be covered by the structural buffer. The additional losses, let's say that we are in the current risk in the red case, these additional losses are going to be covered by the spherical buffer. If, for example, the current losses, the current risk decreases over time, then the severity of the scenario, when we are going to simulate, also is going to decrease. So that means that our new losses are going to be in between the red and the yellow line. So that means that the cyclical component of the buffer is going to be reduced and is going to be moving with respect to the evolution of the risk. Another option that we present here is also considering another case where the reference risk is not the median risk, but it's the minimum risk. In this case, the structural buffer is going to be covering just the losses under the blue line, so under the minimum risk. And the cyclical component is going to be the additional losses that we're going to obtain under, let's say, the current risk. So this red bar on the right-hand side. Let me frame a bit of the paper into the literature. So we have three streams of literature that we're focusing on. So the first stream of literature uses stress as a model to calibrate capital buffers. Then we have, so with respect to this literature, what we do is that we try to find a way through which we can use these approaches, but also understanding which part of the buffer is going to be cyclical and which part is going to be structural. And then we have the nonlinear market economy models. So for example, there are also the contribution by David in this type of literature. So this model, they find that the propagation of the economic shocks changes according to the state of the economy, so let's say recession or normal times. Here we focus on the dimension of the silica risk dimension and we find that the silica risk dimension can affect the propagation of the shock. So we really try to highlight the connection between silica risks and the propagation of the shocks. And then we use this type of outcome from this nonlinear market economy model in the generation of the market economy scenario. Then we have the stream of literature that instead look more at growth risk, so tries to link the relation between, try to find the relation between risk and growth. And so our model also is able to do that and we do that in a multivariate framework that I'm going to show you in a second. So why do we call this market economic model the silica amplifier? We call it like this because the financial cycle amplifies the propagation of shocks, not of all the shocks, but overall, it has a role of amplification of the shocks. And in this case, we're going to focus on the debt service ratio as the main state variable. So in this paper, we focus on the debt service ratio as the variable that is going to provide information and the evolution of level of risks. But we also internally developing also other specification where, for example, we use the systemic risk indicator that is an indicator that we use internally to assess the evolution of silica risks and also we see the role of this type of silica risks in the amplification of the shocks. The model is a multivariate smooth transition rejoin switching model, estimated on a euro area. So what we do is we estimate this model and then we identify the set of macroeconomic shocks. Here I'm going to focus more on the housing shock and the spread shock, but also one could also consider other types of shocks. And also for the implementation of the RIS2 buffer, the use of this structural identification is not a necessary condition. It is more there to help us to understand a bit the propagation and to give also structural interpretation as much as possible to what we find in the linear effect that we find. And we have the state effect. So that means that the model is able to deliver nonlinearity. The propagation of the economic shocks is going to depend on the level of the silica risks. So in this case, by the level of the debt service ratio of the non-financial profit sector in the euro area. So the model is estimated with local projection and it is somehow at the crossroad between the Jorda original contribution where we have like a multivariate model like a VAR estimated those coefficients are estimated with local projection. So for different time horizon and the contribution by the railroad rates where instead we have a nonlinear structure. So that is like the one that we have in equation one where we assume that the economy smoothly transition from one state to another. So it's transition from the upper state to the low state of the economy and the ZT is gonna be our state variable so the debt service ratio was gonna make us transition from one state to the another. So that means that this ZT moves over time and that means also that we're gonna transition smoothly across one state to the other. So that means that we can also have the different risk level. So it's not just the threshold say high risk low risk but it means that I'm gonna have a DSR of a certain level and this gonna be associated to a certain type of amplification. I can increase the debt service ratio in the model of a percentile and I'm gonna obtain a different impulse responses. The coefficient BHU and BHT are gonna be the coefficient used to propagate to obtain the impulse responses from the propagation of the shocks. So the shocks arrive in time T and they're gonna propagate, they're gonna hit the different variables that I have in my vector of endogenous variables and that is with this YT and this is gonna be propagated over time through these matrix coefficient, BHU and BHT. The FZT is this state variable, is this transition variable actually, that makes us transitioning from the high to the low regimen vice versa and it is a logistic transformation of the original state variable, so the original debt service ratio. So it is just a way to obtain a transformation from any way so to have a variable that goes from zero to one. And then we construct confidence interval using the bootstrap approach. The impact matrix is computed from the variance covariance matrix of residual horizon one. So I do the estimation for the different horizon but I focus on order to obtain the shock on the horizon one. So like in a VAR, I'm gonna obtain my variance covariance matrix. I'm gonna do in this case a Cholesky decomposition and I'm gonna project over time this vector, this impact matrix that I obtained. So for the different shocks I have an effect on the different variable and the coefficient BHT and BHU are gonna make me propagate the shock over time for the different variable. The variable FZ there is gonna decide which of the two words dominates. So the upward or the D word. So that means that the impulse responses are gonna be convex combination of these two extreme state of the word. So the high vulnerability case and the low vulnerability case. Here in the paper, what we report is when the logistic transformation is at the 15th percentile or 85th percentile. The endogenous variable are the usual suspects of this type of literature. So we have in this paper real GDP, inflation, unemployment rate, policy rate, the allows prices and the spread between sovereign 10 years bond rate and risk free rate. We chose this variable because they were also the variables that were among the most important variables that are provided in the EBA stress test scenario. So we wanted a bit also to have this type of approach also to think a bit about also in these terms. But of course this type of specification can also evolve according to the different type of demand that we have. The silica risk that we use is the debt service ratio of non-financial private sector. The sample goes from 1999 to 2018. And here I'm gonna report the estimates for the euro area but there are also the estimates for country level. The interesting fact of the country level estimation is that they are also gonna provide as an indication in the euro area where the non-linearity comes from. So let's say they have a strong non-linear propagation on the housing shock. If I go and look at the euro area countries, also I can spot and I can detect where this type of non-linear amplification are gonna be larger. The state variable, as we said, is this debt service ratio. So which is computed as the ratio between an annuity that I pay of my debt D with maturity M and effective lending rate YT. And so I'm gonna compute how much I have to use to repay back on my, how much I have to spend each year to pay back my debt with respect to the income. So it's really this debt service ratio of the whole economy that is gonna tell me how the agents are gonna be vulnerable. So the debt service ratio has this nice feature also that is related to tricky factors of financial vulnerability that is the debt that I have, my capacity to repay. So if, for example, I have a negative shock on YT, I'm gonna have my vulnerability and also the effective lending rate, which is also, as we see nowadays, is also a key variable that can affect the evolution of this variable. Let's give a look at the results. So here we have the result for the housing shock. So here it is, again, it's the same house price shock. So it's a house price shock that arrive in period one. And then the period after is gonna be propagated. So it arrives on house prices, and then the period after is gonna be propagated on the other macro and financial variable. So we interpret this shock as a shock to preference. So the agents want to buy more housing, so this is pretty standard in literature. And what we see is that we're gonna report here the propagation under high vulnerability and low vulnerability. And under high vulnerability, so the red case, the red line, so when we have high risk in the economy, the same shock is gonna be transmitted more to the economy and it's gonna have stronger effect on output, unemployment and policy rate. Whereas instead, when we are in low vulnerability, the same shock is gonna be instead much less significant and less strong, especially in the first two years after the beginning of the shock. I also report here the spread shock. So this is an exogenous increase in spread that also arrives at house prices in the same period. It transmits so you have in the same period an increase in spread and decrease in house prices and then the shock is propagated to the rest of the economy in the following period. And this type of shock here triggers a stronger decreasing output under the case of high risk. Also here, under high risk, we're gonna have a stronger recession. The green line studies the case of low vulnerability in that case, the shock is gonna be less amplified. This type of effect, this type of non-linear effect are also found, for example, in another paper that we have more in the US. What we can see also considering the different economies that the non-linearity on the housing shock is pretty constant across the different type of estimation, whereas instead for the spread shock, it also can depend a bit on the economy that we are considering. Then in the second part of the paper, we try to apply this methodology and to make it a bit more concrete. And we use a very simple and fictional stress test model. So we take the EBA stress test result of 2017, so something really from the stress test exercise, and we try to link the evolution of the city one ratio with respect to the macroeconomic variable. We try different types of specification. In the end, we just wanted to have something very simple, something to make just the point. So in the end, we chose, we went for this very simple equation where we regressed the city one ratio on the GDP, and we find that a decrease in 1% of GDP triggers a decrease of 1% in city one ratio on average. So this was also interestingly a result that was also in line with other stress test model that we were running internally in Bande France. So that's also why we like this type of number. So of course, it's not a number that would take seriously, but it's more to give an illustration and to provide an elasticity between GDP and city one ratio. The Cigar amplifier, so here, is used to generate multiple scenario with the same set of shocks and with different risk levels. And then we use our stress test equation to obtain the bank losses for each risk level. And once we define a reference risk, we can calibrate both the structural component and the silica component. So here, for example, we created a different scenario by considering two very strong recessionary housing shock, so a four standard deviation housing shock and four standard deviation spread shock, that hit the economy at different risk levels. So in the blue line, we have the case of low risk. And in the red line, we have the case of high risk. The yellow risk is the medium risk. So as you can see, the same shock trigger a recession that after two years arrive at 2% in case of low risk. And the same recession instead triggers the same shocks trigger a recession that is three times larger when there is a high risk in the economy. And the intermediate case instead are found for the case of medium risk, yellow line, and the cyclical risk that we assume, of course, it's an assumption being at 75th percentile. So if we use this type of machine, also this type of macroeconomic variables, and we link it to the stress equation that we've seen before, this famous 0.45, we're going to obtain the different shortfalls for the different risk levels. And here, from the risk levels, then we can try to map and obtain rounded levels of buffers. So that's why then we have that, essentially, let us assume that we want to cover with the minimum risk, the median risk is going to be our reference risk. So that means that we're going to take the losses under the 5th percentile, so the yellow losses, to calibrate the structural component of the buffer. So this is going to be the buffer we're going to remain constant across the financial cycle. And let us consider that the risk is at the highest level to the max risk. Then the additional losses are going to be instead used to calibrate the CGR component of the buffer. So if, for example, the risk level then decreases from 100 to 75, we are also going to decrease the CGR component of the buffer, which is going to be just equal to the light red brick, so without the dark red brick. So on that also, I think one point that I also wanted to mention, referring also to David's discussion, so if we look at the different capital stock and according to the different risk level, these type of results seem to suggest that the biggest part of this space should be cyclical, in that the larger part of the losses arrive when there is a higher risk, cyclical risk in the economy. So in this case, if, for example, the minimum risk would be just, let's say, the 50 percentile, then the reference risk would be this low risk, then a big part portion of the model of this stock would be cyclical. So this also goes a bit in the direction that the point was made before. Another point that was also touched in the discussion before was the interaction between capital buffer and borrower-based measures. So what we also claim here is that this type of approach can also help us to think about this type of interaction. So the macroeconomic model that we're focusing on is a macroeconomic model that looks more at the debt service ratio of non-financial private sector. So we're also looking a lot at this type of amplification that arrives on the borrower side. So let us assume that the borrower-based measures are able to reduce the debt service ratio. So then we have also a link between borrower-based measures and the capital measure. So here we do an extreme assumption of borrower-based measures able to limit the debt service ratio from 100 to 75 percentile. Of course, in reality, it is a much smaller effect so far. And so in this case, the current risk would decrease from 100 to 75 percentile. So that means that when, thanks to the action of the borrower-based measures, I can also assess what is the effect on the reduction of the CDR risk. And in this case, so the less severe scenario would imply a cyclical buffer 2 percentage points smaller. Again, the magnitudes are, of course, not realistic, but it's really just a take it as an illustration. So that means that in this way, we can measure how to link. We can measure a link between borrower-based measures and capital ratios. So to conclude, we provide a criterion to jointly calibrate cyclical and structural buffer. We do that by integrating a nonlinear macroeconomic model to the stress test. And the key feature is to have a macroeconomic model in which the higher the risk, the higher the severity of the scenario. And then we can use this also framework to calibrate cyclical and structural buffers. But also, again, as I just showed, we can also think of the relation between borrower-based measures and the capital measures provided that we can assess the effect from borrower-based measures to that service ratio. And also, these risk-to-buffer can also be flexible and adapt to different type of questions, like, for example, the sectoral buffer. So in case we use the model to simulate more certain type of risks that are covered by the sectoral buffer, or also for the positive neutral buffer, in case we want to adapt our partitioning of the macroeconomic space just to the CCYB experience. And then we would have this type of also, then the nonlinear macroeconomic model would help us to do this type of partitioning. So thanks a lot. That's all on my side. Looking forward to discussion. Thank you, Valerio, for the presentation very much on time as well. Very appreciated, too. So I think Marco, not Marco Loduca, but another Marco, it should be online to discuss. Yes. Super. OK. Hello, everyone. I hope you can hear me and see my screen. We hear you and see. Perfect. But I just organized the screen a bit. Alrighty. So first of all, just a quick thanks to the organizers for allowing me to discuss this very nice paper by Cyril and Valerio. So my name is Marco Kos. I'm working at the Monetary and Capital Markets Department at the IMF. OK. And on this first slide, I'm just briefly summarizing what the authors do. I think I can almost actually skip this because it was very nicely explained just a moment ago by Valerio what this is about. So real quick, real brief, this is about the idea to combine a non-linear macro financial model with a stress test model machinery to inform cyclical and structural bank capital buffer requirements. So we've just seen this figure one from the paper, which has these different lines, blue, light, orange, dark, orange corresponding to these capital impacts resulting from different initial conditions, initial levels, different levels of cyclical vulnerability or risk, as the authors call it. Importantly, and that's a point that I will come back to in my comments in a moment, there's this notion of structural risk. And structural risk is defined by the authors as average cyclical risk or average cyclical vulnerability or, alternatively, some lenient quantile of it, so at a level above it. So it basically means it's corresponding to this light orange case, this light orange line, or above it. So somewhere between blue and orange, light orange. And so that structural risk is, and the impacts resulting from this would be used to inform the structural buffer requirements. And then the cyclical scenarios come on top. If this orange current risk, this dark orange line so happens to be below light orange, well then this would imply an additional cyclical buffer requirement. And importantly, and that's also a point that I will come back to in a moment, this capital ratio shift delta car, the shift in the capital ratio for the banking system from the structural scenario is suggested to be used as a backstop. So the overall buffer requirement would be not allowed to fall short of this backstop level. And so overall, I would just say this is of course simply very useful work timely and very relevant work. That's an upfront comment. So on the following two slides, I have two sort of conceptual slides up front before I come to the actual comments and suggestions. So here on the slide, just to kind of recall the basic rationale, the basic philosophy behind cyclical state dependence and RU design, it's basically something that Nelly was also talking about earlier, just like a visual representation of it, that say if we so happen to be in an initial boom phase and we conduct a stress test at this point and we ask what the buffer requirements should be, then we would want to consider a delta shift all the way down to the bottom of the cycle. So this delta would be quite sizable. But this is opposed to a case where on the right side here, let's say we were in an initial recession or close to a recession at the bottom of the cycle. Well, then this additional delta might be much, it should be much smaller to arrive at the bottom. And in a way, like the continuation of these bad conditions at the outset would be bad enough for this cyclical scenario. And so overall, this is just to say that conceptually all of this kind of schematic picture is pretty much in line with various frameworks that are around like the stress capital buffer in the US, the cyclical scenario exercise at the Bank of England, also the ECB, EBA, SSM sort of area-wide stress test exercises kind of, I would say, adhere to this principle, even though there might still be largely like linear models around. But there were cases in the past exercises where some countries, actually Southern European countries, say they so happened to be in this position on the right side of the chart. And then in this case, the shocks that were applied to them were smaller, they were made to be smaller, to not overdo it, so to speak, and to adhere to the cyclical principle. And then growth at risk obviously is a direct nonlinear framework in line with this principle. And even just to recall the sort of very simple methodology that we sometimes also employ at the site, like a two standard deviation benchmark from historical mean kind of benchmark at the site. Even this, if we define it as a deviation from historical mean, which is the horizontal line in this picture, actually does also a result in state-dependent scenarios. And then obviously the author's work is exactly in line with this whole series of examples as well, and in line with this philosophy. So I'm just having one more slide, and then I get to the comments just to complete this discussion and to have an additional perspective on some, say, external shock scenarios, exogenous external shock scenarios, which can also happen. So these can be external demand shocks for exporting countries, supply shocks for import dependent countries. There's FX shock and capital flow considerations here. And the main point, I would say the main takeaway is just the main point I want to make is that in such a case, these external shocks can happen, they do happen independently of the circular position of a country. So these shocks in a way don't care about the circular position of a country. But of course, these external shocks can also be triggering the circular downturns. And that's also a point I may come back to later on. So with this kind of background in the back of our minds, I would come to my comments and suggestions to the paper. I have six of them, six comments. On this first slide, the first point is all around the same notion of releasable buffers and the positive neutral CCYB. Again, first, just to quickly recall that in the paper, Cyril and Valeria define structural risk as the average of cyclical risk or some linear and quantile above it corresponding to times of low vulnerability. And so here is basically a related question then. At one point in the paper, actually, the authors point to this question. They ask it themselves, like how to strike the balance between this so-defined structural risk and cyclical risk. So at what level, basically, to set the quantile, say, for this structural scenario and the implied structural buffers. And related to this, in turn, again, is this notion of the backstop. And the question is, is this backstop meant to be a structural buffer, such as the CCOB, capital conservation buffer that's not releasable and whose use would imply restrictions? So my reading from the paper is that this might be actually what the authors may have in mind, also from the presentation. And if so, then I think the sort of problem would be still the one of, say, no or insufficient releasability of buffers during downturns. So that's my main point, actually, here. And so this reasoning brings me to my first concrete suggestions, actually, for the authors' consideration, which is one that I think it could be useful to consider mapping the whole paper more directly and more explicitly into the positive neutral CCYB discussion. So currently, this is just like one sentence and a footnote in the middle of the paper. So I think it could be useful, really, to expand on this, especially when defining structural risk as average cyclical risk, as the authors do. And then, in turn, related to this in the same context, it would be useful to further elaborate on this notion of the backstop. So do the authors mean structural buffers of the CCOB kind? And then again, a discussion around the releasability of these buffers would be very much useful and warranted, I would say. And again, the overall message here I would find it quite useful to link it, really, to the positive neutral CCYB more explicitly. So then come five additional comments. I'll keep them pretty short. They are on the next two slides. So the first, the second point here is sort of a model-oriented point, tech point about the, say, possible state dependence for bank balance sheets and their response when computing the neutral, cyclical average scenario impacts specifically. So what I mean here is that when employing a stress test model suite to obtain these shifts in capital ratios of the banks from the cyclical average position in the middle of the cycle, then I think it would be useful to consider taking recent historical averages of all the various risk metrics for the banks as a T0 anchor, so to speak. So this would mean default rates, interest rates at the bank level, capital ratios, so essentially the whole balance sheet and sort of profit and loss statement. And why, in turn, well, because I think there might be some additional, some non-linearities which may otherwise distort these estimates of the shift in capital ratios from other cyclical starting points. So it's a bit of a model-oriented detail point, but I was just curious. And I would find it useful if this could be discussed maybe a bit in the paper. The second point, a lighter one, maybe, is about the idea to mention some alternative non-linear models. So on the paper, as we saw, the authors use this smooth transition regime switching model, which is pretty fine. I think it's quite a good choice, actually. I have no particular concerns or so. But it's sort of interesting to see that the authors had to use these sort of sizable shocks for standard deviation shocks to sprites in house prices. Would just be useful to make it a bit more generic to discussion and say some alternative models can come into play, like endogenous and micro-regime switching models, for instance. And I think the reason why it might be useful to do this kind of discussion a bit is also to hint to the fact that such other methods might be more amenable to reflect a circular downturn narrative. So what I mean is that it might actually need small perturbations, small shocks, not forced standard deviation shocks, but small shocks that cause large downturns at the circular peak. And as a quick hint to a kind of theory-oriented paper, let's make this point and showing it in a model, which is also kind of a theory perspective to growth at risk, in fact. One more point in relation to borrower-based policies. So the authors speak a little bit in the paper to this interplay between borrower-based and capital-based micro-roud-assure policies. I should say here that maybe the reason why, let me first read it out and I make a caveat here. So the first point here is to say the authors suggest that borrower-based policies can help reduce current circular risk. When I was reading this, I was thinking, well, may they not also reduce average circular risk beyond the short term? So that's what the authors, again, define as structural risk as per their definition. And maybe it's a bit of a function of the fact that in the version of the paper that I saw, at least, I thought it was not the debt-service ratio. So kind of a flow-flow ratio that Valerio was not talking about in the slides, but in the paper, I thought it was actually a debt-to-GDP ratio, so kind of a stock-to-flow ratio. And that doesn't change my point or affect this point like very fundamentally, but anyways. So my point is this discussion could be maybe extended a little bit and be refined, essentially. To final quick points, in the interest of time, I will not read this out. I guess this is just to say more references to the literature could be established. There's a very useful survey paper by David and co-authors, a recent IMF working paper. On scenario design, let me just mention this. We had a book chapter in the stress test handbook that was published last year with many, many chapters. We have one where we do hint to this idea to use non-linear models, indeed, and account for state dependency and so forth. Then there's a whole bunch of papers. I'm sure the authors are actually aware of these papers. It's just not in the draft yet about the effects of capital buffer releases in general, and especially following the COVID pandemic. And then the last two bullet points are about the positive cycle, neutral, CCYB, the relevant BIS papers would be useful to cite and refer to and discuss. A final point, very short. I think it would be useful to actually consider shortening the papers a bit long. I think the content, it's pretty clear what it is. It's pretty useful, and we'll just benefit from being, you know, constant to less pages, less words. I think it could fit into less than 15 pages, so. So all the references I had are here, and that's it. Thanks very much. Let me pause here. Thanks. That's a lot, Marco, for also taking the time from, I guess, Washington. So let me open the floor. Yeah, we have probably five minutes for questions. You know, David first. What's, Evelyn? Yeah. Okay, so let's go in counterclockwise. So we start here. Thanks. Great presentation. Lovely discussion as well. I just wanted to offer a thought on how to think about the cycle, neutral, structural bit. To my mind, that must link to the costs of raising bank capital requirements. So if you thought those costs were very low, you'd probably want to run the regime with very high structural buffers. Wherever it's high, you do the converse. So I was wondering if you could think about incorporating that, and then that would give you a tied up view about the whole thing. Yeah, my question sort of relates to that a little bit. My comment is that the debt service ratio is a variable that can move around, certainly as we are observing now as a function of what interest rates are doing. So the question is, this framework, wouldn't it be more useful to, rather than looking at the current debt service ratio, to sort of project that forward and consider what would happen to these risks if the debt service ratio were to shoot up as a result of increases in interest rates. But then the question arises that David just asked, to what extent should you prepare for those high readings of the DSR, perhaps even exonetally? And what's the right balance? And how should we think about that? I guess this question also relates to the question of whether the Z variable should the indulgence or what drives the change, which also I think Marco mentioned. Yeah, you had a question as well? So is there a microphone over here? Thanks, Janne Erdanova, Danish Central Bank. I was just wondering, from a more practical perspective, you do have your model on the data going from 1990, 1999 to 2018, and the question we often face doing macro financial policies to a degree, the results are dependent on the historical relationships between the different variables. And what happens to the results when we see a swift shift in the environment as we've currently experienced? Can we still use the results, or do we need to make some adjustments? Because definitely the loss ratios that we've experienced during the financial crisis do not reflect on the environment of high interest rates that we're facing now. And another question regarding the borrower-based measures, I think it's quite interesting to see this interaction between the borrower-based measures and the capital-based measures. But I was just wondering whether the conclusions of the borrower-based measures take into account the fact that it takes quite a lot of time before the borrower-based measures take effect on the stock of the loans. Yes, thanks. Question here, please. Oh, you're okay. Thank you very much. I don't really have any questions. I have a few small suggestions. So, yes, Sar, I think it's a very interesting measure to use for your purposes. But sometimes, in some cases, we know that vulnerabilities can be in disguise of small DSR ratios. For example, DSR ratio of a nation may be small, but a change, a drastic change in DSR ratio could really present vulnerabilities to a financial system. So you could consider that. And another suggestion is kind of address other issues raised by the discussant, including the structural risks and positive neutral CCYB, is perhaps use a variety of different indicators. For example, the output gap could be interesting measure to complement the DSR and perhaps the house price gap, if you wanna talk especially about the structural features of a country or vulnerabilities within the housing market. So that would be interesting to see. Thank you. Okay, I think we need to stop here. I would like to ask you to be concise. Thanks a lot. I'm really surprised by how much both the discussion and the question are really on point on many things that we are also internally starting to work on and to improve. So thanks a lot, Marco, for the discussion. I think it's a very useful discussion. Also, I really like your initial rephrasing of the problem but also in general the point that you make. So I will try a bit to be short indeed. So the absolutely point taken on the fact that the model this type of approach can well fit the positive neutral buffer. It is actually something that we are considering and we are working on that. We are adapting a bit the framework also to consider the differences from the problem that we had there that was more the structural versus the legal. And but it's something that we're trying to evolve in that direction. So thanks a lot. So just to give you also an initial perspective. So this project was started more as you also mentioned for to think about the interaction between the EBA stress test result, the P2G buffer and the country signal buffer. So it was a period where the EBA scenario, there was this idea of having a higher risk for countries with higher severity for countries with higher risk in this type of exercise. And what we try to make in the paper is really this point that we need to understand what we really mean by structural here. So it's absolutely, I would point to what they really like is also this this mentioning of the relation, of the nonlinear relation between the CT1 and the macro variable. So yeah, completely on board with you that it's also a nonlinearity to take into account. And also that's also something where we're trying a bit also to put the nose. Thanks a lot for the suggestion on the other nonlinear models. It's absolutely, I mean, we don't have to be orthodox in choosing one model, but also other models can also help us to achieve this type of severity and risk related severity. For the point, yes, on the board with these measures, okay, I understand the point of trying to be a bit more precise on what are the effects in the channel with respect to the, so what is really structured also there. So it's a dimension that we didn't think about yet. And also thanks really for the practical suggestion about shortening or this type of suggestion that are on super useful feedbacks. For passing to the other points, so the point of David, no, that says the cost of raising buffer also could enter in this type of equation. This is also a very good point. And also this bridges a bit with the positive natural buffer discussion. Also the idea that there is a part of the buffer that even if the risk are not high, since the cost of building the buffer is low, maybe we can converge to that. So it's absolutely a good dimension also to look at, I don't know how much we can incorporate in the model, but it's worth to look at and also to consider more this type of channel, both in terms of new variables or in terms of state dependency. But I agree overall that there is a bit of, the model focuses a lot on this type of amplification on the private sector. So maybe it could be useful since you're also talking of the CT1 buffers to also look a bit more in that type of direction. So it's absolutely a good point. For the use of that service ratio in the monetary policy, also in the general, in the new environment, it is true that it's a nice exercise to think of. And ideally, yeah, that could be also done with an endogenous state variable. So this is not yet implemented, but that could be nice. Also overall, I would say, per se the exercise now of considering how the new environment increase the debt service ratio what are the effects on the amplification also goes a bit in the direction of interaction between monetary policy and the rest of the economy and the financial sector and ultimately also macro proof. So this is also something that it's absolutely a good point and we could actually use it a bit in, for example, in the agile teamwork we're doing where we try also to look a bit in the respect to this type of interaction. So okay, so the practical question, I think it's useful questions also to be precise. So the, yeah, it's true that this is a non-linear model we have with small sample. So it is true the results can a bit change according to the different type of sample and also as I was saying before, also with respect to the country. What I find, what we find in general is that overall there is an amplification of the shocks when there is this high risk in the economy. What can change is what type of shock are more amplified or less amplified. So as I said at the beginning, the housing shock probably is more amplified than other type of shock. Monetary policy shock also is amplified pretty often. But again, for some countries, this is not always the case. Not to mention that including the COVID period completely creates completely different results, but this is a non-linear model. I would be surprised if the opposite were true. Okay, and then I'll wrap up. So, okay, but happy to discuss more later. Thanks a lot. Thanks a lot. No, thanks, thanks. Okay, thanks a lot. Okay, the next presentation in the send is online. So, Luis, before we start, I forgot to say that some one of you lost a room card from the Melia inside Frankfurt stand. So you cannot sleep here tonight. So if you don't have the card, I guess you can ask, I don't know where the baby is outside in the desk. Anyway, if you don't find your room card, it's been found. Okay, so let's proceed with the next paper. So, the others are from the back of the spine. You are connected. Yes, can you hear me? Yeah, I will hear you very well. Can you see my screen? Yeah, well, not now. Okay, let me show it again. I hope it's there. Okay, so go ahead, 25 minutes. So we're over the presentation. So first of all, let me apologize for not being able to attend in person, but our medical conditions really suggested to do this remotely. And today I'm going to be talking about the question whether macroprudential or border-based measures should be targeting non-financial corporations. This is joint work with my colleague Jorge Galán from Banco de España, but of course the usual disclaimer applies. These are our own views. It should not be taken to represent those of Banco de España or the bigger system. So the alarm of the talk is very straightforward. So let me just jump into it. And so, of course, in this venue, it's almost, I don't need even to say this, but the role of non-financial price sector depth in previous crisis has been by now well-established. And this has motivated the introduction of broad macroprudential policies, such as the CCYB. But of course risks might be concentrated in certain sectors, which has also motivated the introduction of targeted macroprudential policies, and those very often have been focused on mortgage debt, for good reasons, since it played a very important role in the last crisis. And the literature has already shown that there is a significant association between the deterioration in credit quality, in mortgage test and default. And by now, already a number of institutions, or jurisdictions, excuse me, in particular in Europe, are already applying household-focused macroprudential policies. However, less attention has been paid to credit non-financial corporations. And in this work, we explore whether targeted macroprudential policies focusing on non-financial corporations can also be effective in reducing the risk of future default of this type of loans, in the same way that they are being used currently for households. And just to motivate this, we see that debt in Europe, in the run-up to the previous crisis, the growth in household credit was actually faster, was actually stronger than growth for non-financial corporations credit. But since then, credit to non-financial corporations in blue has grown faster than credit to households here in green. But if we look at the number of macroprudential policies targeting households, those are much more numerals in Europe than those targeting non-financial corporations. Macroprudential policies targeting non-financial corporations have been relatively infrequent. And if we look at the case of Spain, which is the country in which our empirical analysis will be based, we see that even in the run-up of the previous crisis, while household credit increased substantially, credit to non-financial corporations increased even more here in red. And if we focused on credit to the real estate sector, the increase was really astounding. And those large increases in credit will actually mirror it increases in non-performing low ratios. And while for households, non-performing lows increase substantially, growing above almost 7%, the increase in non-financial corporation non-performing lows was much, much stronger, reaching 20%. And if we focus in particular in construction and real estate, it increased over 35%. And so against this background, the goal of this study is for us to identify the relationship between credit standards, termination and band default risk for non-financial corporations. Our ultimate motivation is to inform and calibrate the operationalization of possible border-based measures targeting corporate lending. And this is not a mere academic question since recently in Spain in particular, we have the authority to implement the needs for corporations based on indebted net measures such as debt to assets, debt to income, the service to income and the interest coverage ratio. But of course, the question really comes to our mind whether we really need border-based measures targeting non-financial corporations. On the one hand, it might seem that the relationship between lending and systemic risk is more clear in the real estate sector. In the real estate sector, there's a clear path for systemic risk since an increase in lending can feed into an increasing prices which in turn might demand more lending to afford these higher prices. And this can lead to a typical and well-established BAMBAS dynamics. However, also in the non-financial corporation, more in general, in the mortgage lending, credit has also been associated with financial crisis and actually the quality of this credit has also been proposed as a measure of systemic risk. And indeed, it has been found that a deterioration in credit quality to non-financial corporations is linked to GDP contractions and crisis. Yet, already now, authorities have a number of specific capital measures targeting the corporate sector, such as sectorial systemic risk buffer or risk-weight add-ons. But, as is the case for household targeted border-based measures, we believe that border-based measures targeting corporations, they do not directly at the bank level, but rather they increase the resilience of borrowers. And for this reason, they can complement capital-based, lender-based measures. But of course, a very important caveat to keep in mind is that if we were to restrict credit to more leverage firms, that could certainly increase the risk of bankruptcies and have a potential negative effect on the economy. And therefore, when we target non-financial corporations, an even higher degree of caution is called for. Nevertheless, in this work, we see this as a first step towards the possible personalization of corporate border-based measures. And what we do here is we empirically analyze the link, the relation between credit standards and the nation and the fault. Because if this exists, then there is a potential rationale to implement these type of measures. If the link we found is not so strong, then the rationale will really be the case. And regarding the literature review, the same important literature now on the markets market, in which on the one hand there have been a number of works showing that there is an important relation between loan quality and areas and repositions. But also there are other works showing that implementing limits based on credit standards can lower excessive credit growth and house prices. And also they can lower the risk of economic constructions and financial crisis. On the non-financial cooperation side, these relationships have been less explored in the literature and they oftentimes have focused on large public-economic firms and one-year-ahead firms. But of course, there is an important literature looking at the determinational and the financial crisis. Determinants of defaults incorporates. We can cite the early work of Altman and Beaver, which again focus mostly on large firms and later on other works, included also market-based information and macro variables, but mostly focused on large cooperation on issuers and oftentimes one-year-ahead defaults. We could also cite another related standard literature, which is the work of Central Bank collateral valuation frameworks. But the very nature of this type of frameworks, they focus on bond issuers, since their goal is to evaluate the quality of the collateral they can use for their operations. And they also focus on short-term defaults, typically one-year, since this is the horizon that the bank would need to sell the collateral in case the partner defaults risk. Also these models tend to be rather complex, and so they would be somewhat complicated to use for established policy limits. Perhaps the closest work to ours is one by Antunes and collaborators, in which they look at a very comprehensive sample of Portuguese firms, including also smaller firms, but they also focus on one-year-ahead default, because their motivation is, again, the Central Bank collateral valuation problems. But still they find that the number of variables linked to indebtedness, such as cash flow and assets ratio, but also interest rate, age, and other variables are important determinants of default. In this work, what we bring in is that we will be looking at a semi-comprehensive firm sample. It will be, for firms in Spain, it will cover a whole financial cycle, which we believe is important to really detect a relationship with risk. And we will focus on bank default over all the credit lifetime of the loan, looking at standards of the nation. This is important for us because if limits to credit standards were implemented, those should be implemented when the credit is granted. And we care about default, not just in the short term, but also in the medium term. And so we believe it is important to consider the whole lifecycle of the loan. So for the data, we rely importantly on the Spanish credit register. This is a very almost comprehensive database of all the exposures of Spanish banks at the monthly frequency. However, it does not have firm information, and for that we rely on the Spanish mercantile register. Mercantile register has information of balance sheet and also profit and loss statement at the yearly level, but it's a bit less comprehensive than the credit register. And we are able to match around half of the exposures. In the left-hand side graph, we see the coverage. There's a slight upward trend, but over all the coverage is around 50%. It will be below 50%. And if we look at sectors, we see that the coverage for large corporations is significantly better than for small business enterprises and real estate, but even for those, the coverage is certainly not negligible. Excuse me. We have a problem in our database and it's that before 2016, the credit register does not identify individual loans, but instead it identifies outstanding credit from a bank to a firm. And so we need first to identify new credits. And for that, we basically look at new bank-firm relations in our database. So relations that were not present the previous month. And also we look at monthly increases in outstanding that are over 10% of the bank-firm exposure because this will correspond to a significant increase in the amount granted. And if there was a limit implemented, it would apply. So it's also important for us. In this way, we identify the new loans and then we follow them forward in time on the credit register and we note whether eventually at any point the loan enters into default. In this way, we obtain a cross-section of new loans within characteristics and an indicator of whether the both ever occur. And since the coverage of our balance sheet data is a bit limited, here we explore whether it gives us a representative sample. On the first graph, we just show the distribution of credit sizes of new credit sizes that we identified between the whole sample in red and the sample with credit with balance sheet information in green. And we see that they look very, very similar. So there are no important differences. But if we look at the percentage of loans that ever entered into default, we see that there are important differences and indeed in the sample with balance sheet information, the defaults are somewhat smaller. If we wait by credit size, then the difference are smaller but still they are there. And this of course raises the possibility of selection bias. It might be that our sample with balance sheet information includes only the best firms. And so the relationship that we identified might not be representative. And this is an issue that we will analyze later in more detail. And let me show you here the relationship, the association between the fraction of loans that ever entered into default and the quintiles of the different standards. So for example, here for the real estate and construction sector, in blue, I show you the fraction of loans that ever entered into default for those in the first quantile in the bottom 20% of the debt to assets ratio. That would be the first point and the last point would correspond to those in the top 20%. So the top quintile of the two assets ratio. And what we can see is that there is a very large increase in the proportion of the faults going from around 6% to over 30%. So this is an increase of more than a factor of five. So it's a really, really large association. If we look at other standards such as the service to income or the to income, the relationship is similar, although it seems to be somewhat unautonic at the top. If we look at interest coverage ratio, the slope is the opposite but this is expected since for interest coverage ratio, larger numbers correspond to lower indebtedness. So again, there's a clear association of in this case, lower indebtedness, lower defaults, high indebtedness, high indebtedness. The situation is very similar for small and mid-sized enterprises and also for large companies. In the case of large companies, if we look at the scale, the dependence is a bit smaller, it's a bit weaker, but still we move from around 3% to over 10%. So again, there's a factor of three difference in defaults depending on the quanta I love, in this case, the type two assets. So here we see a very, very strong association between the standard standard of the nation of future defaults. But of course, this is just association without having any control and we want to be a bit more systematic than that. And for that, our base, the baseline model that we use is a linear probability model in which we regress an indicator variable for whether the low neighbors enter into default over the standard standard generation and a number of parameter long characteristics as well as fixed effects. Long characteristics, we include things like age, size, liquidity, profitability, and collateral. For the interest rate, we don't have the information of the interest rate of each loan, which we would like to include, but we don't have it. Instead, we rely on a proxy, a firm level proxy, which we build as the ratio of interest expenses of the firm over the total depth of the firm over all banks and non-banks. And maybe I should indicate that our baseline variable for default is basically being in a rails for at least 90 days. So here we can see some of the main results. In this case, for small and mid-size enterprises and considering the depth-to-assets ratio, the first column is the regression without any control and in the other columns, we include different controls as well as fixed effects. And what we can see is that there is a very substantial association between in this case, depth-to-assets and defaults. The coefficient is always very significant and it's very robust to the different controls and it's very substantial. We should note that the average default is around 10%. So here, an increase of depth-to-assets from zero to one would lead to a pretty much a doubly in the fraction of the defaults. This is for SMEs. If we look at real estate sector, we find something pretty similar. Again, the association is clear and it's robust to the different controls. Slide is smaller than before, but still quite substantial. And if we look at large companies, again, we find very, very substantial and robust increase, which is always very statistically significant. This is for depth-to-assets. If we look at other standards, income base, such as the depth-to-income, the service-to-income interest-to-assets ratio, which is something similar. Here for the most complete model with full controls and fixed effects, we always find a very substantial and robust coefficient, except the case of the three income for large companies, where the association is not significant, but for other sectors and standards, the association is there and it's larger with the expected sign. So I have argued that this association is very robust, but we have identified a number of variables which significantly affect the association between the standard and the faults. And in particular, we have found that the age of the firm and whether the bank-firm relationship is a new one, it affects substantially decreases the association. So here, column one is just the baseline as before. In column two, we restrict the sample to only firms younger than five years and in column three, we restrict the sample to only new bank-firm relations. And in both cases, the coefficient is significantly lower, decreases between 30 and 50% or 50%. And with the whole sample, but interacting, in this case, the two assets will be indicators for these type of firms. We also find negative and very substantial coefficients. And these holds for the three sectors that we have analyzed and also for other standards. Here, I showed you for the two assets, but the situation is very similar for the service to income, the discovery ratio and the three. These can have potentially important policy implications. Since, for example, young firms on the one hand might be more depending on bank credit in order to grow. And also the prospects might be, the preparing capacity might be more linked to the prospects and to the current balance sheet. And here, what we find is that from a purely statistical point of view, it's also the case that for young firms, the relationship is less strong. And therefore, there is a case to make less stringent limits to these type of young firms. The model so far was linear, but we can of course include more linear coefficients on the different lending standards and they tend to be significant and lead to somewhat concave relations. But this is better appreciated in the graphically. Here, those are the predictive margins, which basically are the predictive probability of the photo according to the model. And we see how these changes when we change the value of the two assets, the two income, the discovery ratio. And this is the model with no linear terms. And what we see is that the lines are almost straight lines. If they were straight, there would be no nonlinear effects. So here we see that the photo assets and the two income, the nonlinear effects are very limited, but for the interest coverage ratio, we do see important linearities. And this might suggest a possible limit for the interest coverage ratio. Since for large interest coverage ratio, we see no effect, but for lower starting for some limit, we start to see a large increase in the probability. And we have performed a number of robustness exercises. Perhaps most important is to check whether our sample is representative or there is the presence of any bias. Since, as I described before, our sample only covers around half of the exposures of Spanish banks. And here we use a Hecman selection model to study whether there is evidence of selection in our sample. And what we find is that the Hecman model, the transverse salts that are almost identical to those of our baseline model. And this is the case for the different standards and also for the different sectors. The results are always very, very close. So we conclude that there's no evidence of selection bias and therefore our sample can be considered representative of the whole universe of Spanish firms. We have also explored including together several credit standards at the same time in the model. And we find that if we include debt to assets and also to get a debt to income, the gain in the model is very limited. But if we include debt to assets and together with interest coverage ratio, the gain time is substantial. So here the different lines correspond to different values of interest coverage ratio. And we see that for the same debt to assets ratio, having a larger interest coverage ratio is linked to lower, clearly lower default prediction. And our baseline model was linear probability, but of course we checked with a logit and probit model and we obtained, here I compare the linear probability with the logit and probit models and the results are almost on top of each other except perhaps for the interest coverage ratio where the linearities are more important but overall the results are very, very robust to the model. And we also include, we explore weaker and stronger notions of default including whether the firm really, whether the bank really writes of the credit or whether we use a subjective statement from the bank indicating that the credit is dubious and the results are very, very similar. We compare analyzing bank and non-bank debt separately and we find that they both have strong associations but the association of bank debt is stronger. And if we look at different sub-samples, we also find association in every case although it's somewhat weaker post-2011. And with this I reached my conclassions. We have find a very strong and robust relation between credit standards, bank trugination and bank default frequency. It's stronger for debt-to-assets and interest coverage ratio and especially for SMEs and real estate firms but also for the other sectors. The relationship phone is weaker for young firms and new bank-front relations. Considence very standard simultaneously adds discriminatory power and so there is some scope for complementarities of using different, potentially different measures at the same time. And we hope this is our first step towards the calibration of macroprudential border where based measures track the movement of corporations. This is promising, showing that there is an important association but of course care must be taken and we need to be mindful of the possible negative effects. And with this I welcome your comments and questions. Thank you. Thank you very much for the presentation. So, yeah. So now the other Marco will do the second discussion. Good evening. Thank you for the invitation and it has been a pleasure actually to read this paper, it's very clear. And a few more minutes of energy before concluding I will try to not take too much time. So let me briefly summarize again the key points of the study but very briefly. So they find a link between landing standards of originations and the fault risk on corporate loans. This link is different depending on the business type of the firms and it's very robust across different tests that they do. And they rely on a very nice data sets, is micro data for like the exposures they use the credit registry and then they match essentially with the firm balance sheet data gain from Bank of Spain. So this paper contributes to the discussion on borrower based measure for the corporate sectors by showing obviously the importance of landing standards at origination that of course could be a variable that policy makers can control. And also because it offers some insights on the possible calibrations of the borrower based measures especially on the role of firm heterogeneity in setting borrower based measures. So I have a few comments, a few technical comments and then some general points for a broader discussion on borrower based measures. Let me start with the technical comments. You see, or you saw the equation that is the econometrics, the linear probability model. There is the default indicator when a company default on an exposure to a bank. And this is explained by the landing standards controls for firms that are firm and exposure specific. And then the fixed effects by bank, sector, location and time. Now, I must say that here the devil is really in the details because there are, there is a lot of things that you need to do to construct these variables and the paper is really clear. But for example, they can approximately identify new loans as we saw from the presentation. But then they cannot identify on which loan the default occurs. So they assume the default on the wall exposure. So the paper is very clear about this. They do robustness and of course there are pros and cons. But let me go to my suggestions. The first one, maybe the authors will not agree but I'm convinced that there must be some nonlinearity in this data. I really expect that the relationship between the credit standards and the fault is nonlinear. Of course, they explore some nonlinearity, they especially a quadratic form. But I will try to use the buckets that they also use at the beginning in the regressions or explore models with threshold effects. Why? Because the discussion of thresholds is very important for the calibration of border-based measures. And I think that I would like to see a bit more on that to be convinced that there's really not much nonlinearity in the data. And I think some nonlinearity is there. Maybe not for all the categories that they analyze, but for some, I think for some categories of further, you can see that there is something. The second point on nonlinearity is the interruptions among credit standards. So what if two credit standards are high risk? Does it amplify the effect on the default risk? There is something there, but I think even there, there could be more. And again, the interaction with the buckets could be interesting. Another point, OK, it was not in the presentations, but essentially all the controls are locked. This is to use information that is available in real time. But I don't think it's strictly needed in this type of paper. In this paper, they are not assessing whether variables that are available in real time have forecasting power. But they are after the true link between credit standards and the fault risk. So if possible, and if available, I will use contemporaneous variables. One point on the time horizon analysis. So the horizon of the fault is essentially indefinite. They look at whether a loan has ever defaulted, which is good, but there is one detail that I didn't find in the paper. And I would like to know whether the loans that are currently performing are excluded from the analysis. Because of course, these bring a bias. You don't know what will be the future of this loan. They can default in the future. And again, also to maybe solve this issue, maybe they can run some robustness on specific time horizon. So of course, focusing on the long term and focusing on the credit origination, but setting a time horizon of five years on 10 years and see whether the results hold. And one last point, as I suggested, I will focus as a baseline on the joint estimation on the model of the joint estimation on the lending standards. This is because in this way, you can assess the joint significance. And I think that there are no constraints in the sample. There are 100,000 observations, so you can do that. So maybe these are some suggestions. And maybe here are one chart to make the point that nonlinearity is important and also the interactions, especially in mortgage markets. This is what we find for the LTI. So we see the default rates for mortgage loans by quintile of the LTI in blue. And then you see that if you condition this to a high LTV too, then, and we have the yellow bar, then the default become higher. And the fact is concentrated in the top, in the fifth quantile, essentially. So there is interaction and nonlinearity. And there is, especially, we see it for the mortgage loans. A couple of ideas that I had when reading the papers. This is not really for this paper, but for an expansion. Of course, in the results that were explained and every coefficient is really explained why it has this sign and not the other sign. There is a robust result on the role of liquidity in lowering the fold risk. And of course, this can open up the issue of whether we want also some sort of liquidity standards, at least for certain type of firms. And finally, I think there is a lot in the fixed effects of these models. For example, the bank fixed effects tells you whether a specific bank underestimates systematically the risk of default. And then, of course, you can link it to best bank specific features. I think this is quite interesting. And I didn't see it explored also in other papers. And then, of course, you can even interact with time to see how these bias evolve over the cycle. So a few ideas. Now, let me come to my comments in general on the borrower-based measures. I will say that borrower-based measures for mortgage loans are already a bit of an headache for policymakers. Forget about the political implications. But of course, it seems an easy setting. You just have two players, the lender, which is the bank, and the borrower, that is normally the household. There is a relatively simple enforcement. The banks and the supervisors take care. But we have country specificities. There is multiple choices of instruments. There are different business types. Think about first-time buyers versus the buy-to-let. And this essentially results in already heterogeneous approaches to the setting of borrower-based measures. And here we have an example of the dispersion in the euro area. This focuses just on policies for first-time buyers. And we see the LTV ranges from 80% to 100%. And we know today who is the 100%. And then the DSDI ranges from 30% to 80%. Here there is a little detail, which is the definition of income. Of course, that matters a lot. And then also the exemptions varies a lot across countries. The exemptions is the fraction of loans that can deviate from the central limit. And you see, for the LTV, they deviate from 10% of the total loans to 35%. So even in this simple setting, there is a lot of margin before changing the calibration. So the key question then is, how would borrower-based measures for corporate look like? So here we have several players. The lenders are banks and markets. The borrowers are firms with different business types. And of course, the message is the calibration, one calibration cannot fit all. We already saw from the results of the paper that small-medium enterprises, young firms, large corporations, they probably need different calibrations. And probably there is in general a role from the dependence on external finance in setting these measures. But then the question is, really, how do you design them? Do you set different limits across business? Do you set the central limits? And then you set exemptions for different business? Or you just focus to simplify on a systemic sector, for example, the real estate? And then, of course, there is the issue that enforcement is challenging because firms can get funds from banks and markets. And I'm sure there could be leakages. They will find their way through. And I think we have already some evidence of leakages. We worked hard on that. And here it is. I think it's uncontroversial. And I think that's it for me. Thanks a lot for your attention. Thank you, Juan. A very, very impressive discussion. So I think it's time for dinner for a couple of questions and then we could flow back to, yeah. So let's start here then, Adelaide. Thank you very much on a very interesting presentation, such as an important aspect. And I think what the authors find that underwriting standards are really important for default. That is no doubt. And I was wondering if authors have really thought about using forward-looking information on a given enterprise or a project. Because most of the banks, I think, when they evaluate company credit worthiness, they are looking not only at backward-looking information like debt-to-asset ratios and whatnot, but also forward-looking information in a particular project, whether it would generate sufficient capacity in terms of money flow. Another thing, I didn't really catch it. Perhaps the author could clarify whether they are using a loan default model versus firm default or company default model. Because in many cases, companies do have many loans in different banking institutions. And for example, they may have trucking leases for 100 trucks and whatnot. So there is a bit of a correlation across these defaults when you consider a single company. So yeah, at least in my experience, we did similar models. And to address that, we did firm defaults. And another thing that's important not also for the estimations, but also for implementation of potential macro-predential policies, whether these financial ratios and financial statistics are based on a company level or group level consolidated. Because in many times, it may get a subsidiary company, which is backed by a mother company operating within a group. They're within firm, across firm different relationships. And sometimes you have a guarantor, a personal guarantor, like a household or a guarantor in terms of another company. So that's also one thing we have to have in mind when both estimating and implementing macro-pru. And I'm not sure if you've used, but many authors also use the number of relationship with banks and the length in terms of time of a relationship with a specific bank, because it may also be kind of a good predictor of whether a firm could default. Thank you. OK. My question goes to the heterogeneity across sectors. I understand that this is one of the start-up reasons why it may be quite difficult to implement borrow-based tools for the corporate sector that corporate finance tells you that, and I think Marco alluded to that, that different types of businesses may have different debt-carrying capacity. And I think, actually, you could look at this in the context of your paper, because I think that sits in your fixed effects. You have fixed effects for subsectors. And you could examine whether these fixed effects are significantly different. You could look at the distribution of these fixed effects and see whether there's a whole range, how they distribute to see whether there is indeed this heterogeneity that could make our lives more difficult. Thank you. Two questions here. Then we close that with our hand. Sorry, we don't overlap. Let's start here. Then we come back here. Two questions. Thank you for the presentation. It's a very interesting paper. And I'm wondering if it's not a question, but a suggestion more, if using this very rich database that is very gives you the possibility of deepening the analysis of the differences between the companies, could be, for example, separate SMEs from other large companies. But maybe could be analyzed other characteristics that could explain the behavior of the company compared to the average of the sector. I mean, using standard classifications of sectors, of activity, and using ratios from the balance sheet compared to the average of that ratio of the sector. And then looking at if some of these differences of the ratios compared to the average have relevance in the mortgage, not mortgage, or in the credit in the defaults or the credit worthiness of these companies, how they behave. It's kind of different, different. Well, it's probably what you mean. One question there. I really like the research question. I just wonder. So you say that lending standard have an effect on default. This is no doubt about that. But I think one missing link there is that I wonder if you could explore in your data the link between macro potential policies that leading to those changes in lending standards. So maybe buying chain lending standards because of their own risk appetite changing. But it could also be that during the time period you observe changes in CCYB policies or LTV policies. And then if you could link that to the changes in the lending standard in your model, then maybe the link between macro potential policy leading to the change in default would be, I guess, more clear. So it's just a suggestion. OK. Thank you. OK. We'll be cut and then we come back here. Maybe let's proceed here for a reason of practicality. Thanks a lot. Katarzyna Wodnik, European Central Bank. So two questions as some of the questions I had were anticipated already. So two are left. One is very much along what Mark already said. I was wondering, watching your presentation, why didn't you go for something like a survival analysis? And I tell you why I'm bothered. Because the way I understand your analysis, it's like you pull together loans which are working capital loans and investment loans. And the policy message we give is that you impose more or less the same borrow-based measure on both irrespectively of how long they are supposed to work. And whereas I would say that when a bank gives a loan for one year, it should take a bit much less of a risk than for five years. And I would gladly distinguish it. And survival analysis could probably work with your data to a degree. I understand that using more without losing many observations. And the other thing is like a more conceptual. So I agree about everything set about sectors, but there's something broader than that. So when we talk about households, that they should have some borrow-based measures goes without saying. When I think of firms and I think about debt to assets, what I remember from accounting classes is that the higher debt, the lower taxes. And this means that to a degree corporates, in contrast to households, benefit from leverage. And that is why I would expect that non-linearities, which we observe for households, will be different than non-linearities we expect for corporates. Or maybe the relationship will appear in different parts of the distribution. Some discussion of it, some exploration, not copy-pasting from households to corporates would be really evaluated. Because what you touched upon is really inspiring. So super work. Thanks a lot. OK, last but not least. Hello, Lars Speckman from Orus University and SAFE. I guess I just wanted to echo the point that you got that don't just throw everything into the fixed effect, but rather try to explore them a little bit more. And what I was thinking was to explore the time dimension, maybe interacted with the credit standards to see if there's something over time. But as a general point, sector and time heterogeneity is important for macro-planetary policy. So just don't put them all in the fixed effect. Thank you. Thank you. So, Luis, I know it's a cliche, but you really literally stand between us and the finest German cuisine. So it jokes apart if you can be concise, please. But you got a lot of comments. But perhaps you can wrap up the most important ones. Thank you. Yes, I do my best. First, thank you all. Thank you, Marco, and thank you all the commenters. They are very, very good points. And just a few points on the heterogeneity part and on the linearity. We do see some nonlinear effects, mostly between the two assets and interscorporate ratio are mostly for arts corporations. But for others, we see less than that. I take your point. We can explore this in more detail. We can expand looking at thresholds and doing this even in different ways, perhaps non-parametrically. But here, this is a good point. On the fixed effect, as several of you commented, we do have sector fixed effects at the two-digit NAIC code. And we haven't looked really, you are right, we haven't looked at how average default varies across sectors and this is something we can explore. Perhaps most importantly for us would be whether those sectors respond differently to the assets or to the different standards. We see that across the board, three sectors reconsider that there are important differences and perhaps this is certainly worth exploring whether we see an even more fine-grained difference. So in the model, we do include also performing loans. So we include those that perform and those that don't. We include all of them. We do exclude those that atrogenation are already with some level of default to avoid having zombie lending. Although this has a small effect because there are very few of these type of credits. And we do explore the effect over time of the standards. And we find that the effect is higher during crisis times. So apparently during crisis times, it's really when these measures are more informative. And we find that in more recent times, the effect is slightly smaller, but still is there and still is pretty significant. We can look at the effect of different macro-potential policies in Spain. So far, there would be mostly capital bumpers which might have affected some banks and perhaps they have strengthened, they have limited some of the standards. And we can see the effect of this. Some of you suggested. The rapid analysis is something we have considered. We do sometimes we have explored looking only at default in three-year or five-year horizons. And the results are pretty similar. But we can do this more systematically with a duration model, survival model, as you suggest. And about groups, we look at the individual firm level, but we include a dummy for whether the firm is part of a group. And we find that this dummy is important, especially for smaller firms. So firms that are part of a bigger group default less. And this is not too surprising. For larger firms, it has much less importance. And we find that the effect of the two assets or the other standards is not affected by whether the firm is part of a group. And well, I think I will leave it here. Just thank you again for the suggestions and comments. They are very well taken. Thank you. Let us feel if we don't have applause for the session.