Thank you. Good morning, everybody. Great to be with you and great to have an opportunity to talk to you about this work. This is work that I have done with colleagues at Harvard and a PhD student at the time, who's now a professor at Bocconi. It's called Predictable Financial Crises. This is work that originated out of really our reaction to the central narrative out of the 2008 financial crisis that went around the world. I have here three quotes. These are from three of the leading policymakers in the United States. In fact we had an event at Harvard in 2018 where we brought some of these folks back together again, and really trying to understand the drivers of financial crises at the time and how much they were potentially preventable. What you can see here, I start with this quote by Tim Geithner, who says that 'Financial crises can't be reliably anticipated or pre-empted.' Hank Paulson, the Treasury Secretary at the time in the United States, 'My strong belief is that crises are unpredictable in terms of cause or timing or the severity when they hit.' Then Ben Bernanke with a similar quote, saying, 'The crisis involved a 21st century electronic panic by institutions. It was an old-fashioned run in new clothes.' Now, I think the central message from all of these quotes is that financial crises are essentially unpredictable. I think if we look back, at least my assessment of the US policy response - and you can quibble with it - but was that once things got going they sprang into action very quickly. So, in the sense that they were really very good at firefighting. To us, the central question underlying this work was, what about policing things beforehand? Was there a role in how predictable are financial crises before you get to the point where the house really feels like it's burning down? Now, you might say, 'Robin, isn't that a question that people in finance and people in economics and academics have been looking at for decades? What possibly could you bring to this question? Isn't this the kind of thing that the IMF studies every day? Isn't this the kind of thing that we have a pretty good handle on?' To be frank, we thought the answer to that was yes. But then when we started looking at it, one of the things we realised was that in fact, while there was a recognition in the policy community that crises were preceded by weak economic fundamentals, if you think for example some of the work by Carmen Reinhart and Ken Rogoff - you might know some of that book or some of that work or maybe the book by Reinhart and Rogoff - there was still the general belief that crises were largely unpredictable. This really struck us as being different from, I would say, an emerging alternative. Actually, I would call it the Wall Street alternative, to be frank, which was that markets had a Minsky Moment, that there were credit bubbles that became undone, and that when the credit bubble came undone, that was sowing the seeds of the next crisis. Now, to be fair, there is work that does that. In fact, there are great pieces of work at the Bank for International Settlements - at the BIS - where people have been looking at these issues. But I think it was largely under the radar and also didn't really give you simple numbers for how predictable you would think of a financial crisis as being. So, what do we do? In this work we actually put together a historical database of financial crises from 42 countries from the 1950s through 2016. In fact, we're not reinventing the wheel here at all. We are just using data from conventional sources. We are using historical data on the growth of credit to businesses and households to try to forecast financial crisis. I would say the central concept underlying our work is that financial crises are the by-product of a credit bubble. Now, what is a credit bubble? A credit bubble - different from an equity market bubble and different from a stock market bubble - credit bubbles are quiet. They are during periods of calm where nothing is going wrong. Businesses are not going bust. Households are not going bust. They're different from equity market bubbles. Equity market bubbles are noisy, there's lots of trading volume and speculation and people getting rich. Credit bubbles are again, things are quiet but booming. We try to capture this idea of a credit bubble using the combination of two factors. The first is that you have a period of large credit growth. Then the second is that is occurring concurrent with growth in asset prices. We do that in two settings: in the household sector and we do it in the corporate sector. Let me just give you - make this concrete for you in the household sector. The idea is when housing credit is expanding and house prices are going up at the same time. In the corporate sector it would be companies are borrowing a lot and markets are up. When you have that combination of things, that is a marker of an enhanced supply of credit. We use that, we construct a variable based on this and we use that to predict financial crises. We find actually that the combination of these two variables is remarkable at predicting crises. When credit markets are overheated in the sense that we define it, the probability of a crisis rises from about five to eight per cent over a three-year window to about forty per cent. You might also ask, 'How much lead time do you have based on an indicator like this?' In other words, if I am producing weather forecasts and I'm only able to tell you about the rain five minutes before it happens, you didn't have time to get the umbrella on the way out the door. What we find is that you have about a two-to-three-year horizon for forecasting crises. In fact, it's much easier to forecast in a two-to-three-year horizon than it is at a one-year horizon. It's actually like that in the equity bubble space as well. If you're really trying to characterise the collapse of a bubble, it's really hard to do that at a short horizon. At a longer horizon you can say much more. So, let me get into it. I'm just going to summarise what we do. The first thing that I'm going to show you is that if you just do simple things like look at the growth of credit alone, it's actually quite hard to forecast crises. Really you need this combination of credit growth and asset price growth at the same time. We construct this variable. We call it the red zone. In fact, when we first were doing this work we called it the J zone, after Jacob, who was the co-author who was really doing a lot of the work and the data work. But then ultimately we decided the red zone - for reasons that I'll get into. The second thing that we discover is - and I found this very surprising - that the overheating in business sector and household credit were really quite separate things. I thought of those as maybe happening at the same time. But in fact if you look for example in 2005, 2006 that was largely a household-led credit bubble. But if you look at other points in history and other countries, in fact they're largely corporate events. I guess I'll ask the audience, maybe ask Michael here, when was the last time we had a mega credit bubble that involved both, and which country? Any ideas? US? No. This is one of the most spectacular. I always like to remind people of this because it was the most spectacular credit bubble in modern history; it was Japan in 1989. That was an episode where you really had all things on fire, real estate on fire, household borrowing, corporate borrowing, asset prices through the roof and so on. Of course Japan's been living with that hangover for a very, very long time. But again that's actually the exception. In general if you look at these episodes in history, they are somewhat separable. It's not to say that they're completely uncorrelated, but they're much less correlated than I expected. Two more findings. Number one, there's a global component to this. Of course if I detect a credit bubble in country X, that is useful for forecasting a crisis in country X but it's also useful for forecasting a crisis in country Y - even if country Y doesn't look quite the same. In other words, there's a global component to credit bubbles. I'm going to show you how that looks in the data. The easiest way to explain this is to say, Germany technically wasn't really in a credit bubble in 2005. But if you looked around, it was sort of surrounded by countries that were looking like they were in credit bubbles. That told you that Germany was in a precarious situation as well. Then the last part of the paper here I'll spend the least time, but this was our attempt to provide advice to policymakers. The real question we're asking here is, how high should the probability of a financial crisis be allowed to climb before policymakers should do anything about it? Here's how you think about this. I'm just going to show you that the probability of a crisis grows as high as 40 per cent, even 45 per cent. Now, you might say, 'Okay, Robin, you're patting yourself on the back. Forty-five per cent probability of a crisis, that means fifty-five per cent probability you're wrong. You're more likely wrong than you're right.' I would say yes, but if you have a warning of something that could happen and the thing that will happen is quite dangerous, maybe that will cause you to take some steps. So, in the paper we do a lot of work to try to figure out, exactly what should that balance be? How aggressive should policymakers be in response to these potential threats? One of the expressions that my co-author loves to use is, he says, 'When you're getting really close to a crisis, that's the moment when you want to pad the runway,' right? That's one, or if you're further away - you're two or three years back from a potential crisis, maybe you're a booming credit market - those are moments where maybe you put on the brakes a little bit. Slightly different types of ways of thinking about what a policymaker should do. I'm going to skip some of the previous literature. I would just point you to the most famous pieces of prior work here, which are of course Minsky and Kindleberger. Kindleberger, who wrote the definitive and best book on crises I think in history - if you haven't read, it should be a required reading for anybody interested in the history of financial markets - and of course Minsky in his treatment of these issues as well. All right, let me tell you about the data. We use an unbalanced panel dataset covering 42 countries from 1950 through 2016. In fact, you can go back much further but 1950 through 2016 is the period that overlaps with where we can find data also on credit growth and asset price growth for these countries. If you just want to have a database of financial crisis, actually you can go back even to the 1800s, as many of you know. Our key dependent variable, the thing we're trying to forecast, is literally just a one or a zero. It's a one if you're in a crisis and it's a zero if you're outside of a crisis. This is an indicator variable that we get from Baron, Verner and Xiong. There are two other famous datasets on the history of financial crises. Reinhart and Rogoff is one of them, and then Jorda, Schularick and Taylor. We've looked at these other datasets. They give you basically the same sets of results. What are the things that we're using to forecast crises? This is not actually some big data exercise. It's actually everything that we do you could fit into a pretty small spreadsheet into Excel. Two variables: the first is the growth in credit. Here we're just computing credit to GDP in every single year, and then we're looking at the three-year change in that. That is a variable that other folks have used in the past as well. We do that separately for businesses and households. Then the second variable we're looking at is we're pairing that with asset price growth in that market. For businesses we're looking at asset price growth in the stock market. For households we're looking at the growth in housing prices. The first one is from global financial data, the second one is from the Bank of International Settlements. As you can see, really very straightforward pieces of underlying data here. So, what is a crisis? It's a question you might ask. We kind of know it when we see it, right? We were in one in 2008. We certainly think that Japan was in one in 1990. We're using their definition from this dataset that we're using, but what is their definition? Their definition is a combination of three things. Bank stocks falling by 30 per cent or more in that year, widespread banking failures, and severe withdrawals from banks. If we're finding the combination of those things then they will classify that year as being in a crisis. I just want to be very clear. This is different from an equity crash, right? For example, you would say the year 2000 in the United States and in many other markets was an equity market crash. It was not a banking crisis per se. We're really looking at widespread crises, the kind that cause I would call it economic devastation historically. The unconditional probability in the dataset of a crisis in a given year is four per cent. Translation: every 25 years happens for the typical economy that you see in the data. We're trying to see if we can do better than that. Of course it means that if you're just flipping a coin, essentially, every 25 years you should get one. We want to try to see, are there moments when that probability is elevated greater than that that four per cent? The first thing that we do is we try to forecast the probability of a crisis using just credit growth. I will skip the details of this regression. But the reason we did this is because of prior work that uses just credit growth. One of the things that you find here - and you can see this if you look at the R-Squared numbers below and the predictor variables here are changes in private credit, changes in business credit and changes in household credit - you can see that the R-Squared numbers are tiny, often one per cent or less. In fact, it turns out that predicting a crisis to happen, if all that you have is the credit growth, works - it's statistically significant - but the effects are quite small. Now, here is where we try to add some value. This is again a very, very simple cut of the data and all we're doing is sorting the data into two dimensions. From left to right here we are going - and I would have you focus on the very top, that top set of numbers, so the top left. From left to right, we are going from low credit growth to high credit growth. Then from top to bottom we are going from low asset price growth to high asset price growth. In other words, if you go to the bottom right, that is the periods where you have very high credit growth and very high asset price growth. What are each of those numbers telling you? They are telling you the probability of a crisis over a three-year window. You can see here that for example if you go in the top left, the probability of a crisis in a three-year window is just 4.2 per cent, right, which is really close to the unconditional number. When you move to the bottom right, that number rises to 45 per cent over a three-year window. The right-hand panel is just literally the same thing as that left-hand panel, but it's just differencing everything from that centre cell. Do you see that centre cell where it says eight? We're literally just differencing from that cell so that you can see what's the incremental effect relative to the baseline. The bottom panel shows you the same analysis using household debt and house prices. Here you can see the same effect, which is to say that in that bottom-right corner the probability of a crisis when you have a substantial credit growth compared with substantial asset price growth, the probability rises to 37 per cent. Now, this is a simple bucket-based way to understand what is a complex multivariate relationship interaction. If this were a large dataset presentation I would have said this was a machine learning model that was capturing the non-linearities. It's easy to see what the non-linearities are here. It's really in that situation where you have a massive combination of factors. That is where probability of a crisis is elevated. We call this, that bucket that I've highlighted, the red zone. We call it the R zone. All right, so let me tell you about the red zone. The red zone, as I said, is when you have high debt growth and high price growth or high asset price growth. Now you can put this into a regression to try to estimate what is the probability conditional on the R zone of an increased crisis. I already showed you the main number, so really now I'm just doing it in a slightly different way, in a regression-based way. If you do that in a regression-based way it looks something like this. The reason that I wanted to show it to you in a regression-based way is, you can see the difference at different horizons. If you look in that first set of columns, if you're trying to predict a crisis one year out, even with this killer variable it's actually quite hard. In fact, the probability of crisis conditional on the R zone, it only goes up by about five percentage points. Once you're looking two or three years out, the probability is substantially elevated. You can see that in the column that is labelled crisis within three years and that I've highlighted. You can see that there, we're starting to really have a bit more bang for your buck. Again, as you can see, if you just use that variable you have a probability of about 34 per cent in addition to the constant term. You get the same thing if you're looking at the household sector. So, what I was just showing you was for the business sector. If you do the household sector, again somewhat weak at a one-year horizon. At a three-year horizon, a substantial amount of predictability. Those are the main things that I want you to take away. But I think if you're looking at a presentation like this, you probably have a bunch of other questions. I want to just tell you, we had those questions, too. I'll take you through our responses to some of those questions. So, one question you might have is, this is one dataset, it's one set of historical episodes. How robust is this, and what might the robustness questions be? Well, is this driven by the financial crisis of 2008? What if you split the sample? What if you define things differently? What if you define a crisis differently; do you get different results? Does that change our conclusion? Our overall take on this is essentially no. That said, this is an ex-post historical analysis and you have to ask yourself always, whenever you're doing an exercise like this, you have to take the data with a grain of salt. I'll say something that's not on the slide but that we worry about, that many people have asked me is, they've said, 'Well, now that we know all this stuff, will it not be true in the future?' I would say that would be success, right? Or a related critique that people say is, 'Well, if we had a policy response and we understood what was happening at the time, does that change how we think about things?' I would say, yes, it does change how you think about things but it's a little bit depressing if we thought that we understood this and we still had the level of predictability that we do in the data. Another set of questions you might ask is, how correlated are household credit booms and business credit booms? I'll show you that. How much is a global factor? Then again, what should people actually try to do about all this? So, the first thing I want to show you is business versus household credit. Overheating in business and household credit markets are separate phenomena. A simple way to do this is actually just to put both of them into a regression to try to forecast a crisis. If you put them both into a regression, you find that they both work in about the same magnitude if we're forecasting a crisis. The shocking thing to me is that the correlation between a business credit boom and a household credit boom is only 16 per cent. As I said, it's actually very rare but it does happen, like Japan in 1988, 1989, when they were occurring in tandem. When that happens, when you're having both at the same time you can see this number up here, 65.4. That is the probability of a crisis when you have both a business credit boom and a household credit boom. That is a killer indicator, but it happens so rarely in history that it's not that helpful. In fact, we'd like to have something that turns on a little bit more often. The second set of questions you might ask is local versus global overheating. What I'm showing you here is the percentage of the countries in our dataset that are either in a business credit boom or a household credit boom. Now, you can see here there certainly is a global component to this. To me, the other interesting thing when I look at this is that once you aggregate across countries, they do seem to be correlated. But it's a little bit misleading because the timing of the household credit booms and business credit booms is actually not very synchronised. Again, this is the risk of doing what I would call ocular econometrics of really where your eyes are fooling you a little bit. Some of these time series are mistimed or not synchronised by a factor of two or three years. In any case, if you were to try to do the exercise where you say, 'Let me use my local credit boom as well as the global factor to try to forecast a crisis,' you can do that here. Here, I've put both of them into the regression. Again, I'm trying to forecast the probability of a crisis within three years. You can see that both are actually quite helpful for forecasting a crisis. Again, think of it as Germany might not be in a credit boom by itself but if the neighbouring countries are, that's a useful piece of information for you for forecasting purposes. The last set of things that I want to talk about - and I won't go through all the maths on this - but I want to just seed the idea with you a little bit, which is that while this red zone has substantial predictive power for the arrival of a crisis, of course it's generating false alarms. Then sometimes it's wrong, so how do you deal with that when you're having a forecasting problem like this? How strong should the predictability be for you to do anything about it? Now, it turns out that you can construct this notion of an R zone in very different ways. For example, if I wanted an indicator that told me - that turned on at the slightest probability of a crisis, I could design an indicator like that. Of course if I had an indicator like that, it would be turning on all the time. If I did something about it all the time, I would be often acting on imperfect information. On the flip side, imagine I could turn on an indicator where it was right almost every time it came on. So, it only turned on in the most extreme circumstances, like in Japan in 1988, when everything was completely on fire. It could do something like that, too. The problem with an indicator like that is that I'm often going to miss a lot of stuff because I'm not going to turn on when there are actually other things going on that are actually quite risky. You can see it's difficult. When you're trying to put something together that is going to be useful and not turn on too much, not turn on too little, a difficult problem. We ended up looking at this. We had stumbled on this zone, this R zone, where we thought it was an interesting area but thought, well, maybe actually it's too strict. Maybe we should have an indicator that actually turns on more often and is wrong more often. Maybe we're okay being wrong more often because the costs of missing a crisis are really, really high. I'd rather do something about something that could've happened than avoid doing something about potential action that I missed. So, we use the data to construct - this is getting very technical - but what we call a policy possibility frontier, which is essentially, think about that trade-off of the cost of acting when you're wrong versus the cost of missing it when you're right. To do that, we really have to think about a contingency table, which is a representation of the predictive efficacy of an R zone indicator. Here we're looking at the four types of things happening, the true positives; that's when I'm thinking there's a crisis and there is. The true negatives; that's when I think there's no crisis and there isn't. The false positives, which is that my indicator says yes, something will happen and it doesn't. The false negatives, when the indicator says that nothing will happen but it does. It comes down to really thinking about the trade-offs between these four types of numbers. Again, there are some technical aspects here. You can compute something called the positive predictive value and the negative predictive value. But these are really guides to how you would use these numbers. I'm going to skip through some of this. I want to just show you this last picture, which I think gives you a good sense of some of these trade-offs. This picture, I'd like to end the paper on this, because it's really a summary of our entire dataset. On the vertical axis we have price growth and on the horizontal axis we have credit growth. Each of those triangles or circles is a crisis. What do you see? You see our red zone, first of all. The red zone is, most crises happen following massive credit booms in the household or business sector. Still, you can expand that to what we call a yellow zone. If you expand it to a yellow zone you will capture an even larger fraction of those potential crises. Of course, even there it's still a model, you might miss some - and you see that we missed a handful of them, like Turkey in 1994 and Germany in 2008, although Germany in 2008, I kept bringing that up because if you were to look at the neighbouring countries, of course you still would've caught it. As I said, we talked about the policy trade-off but I'm out of time. I did want to show you, I think, one of the big questions, which is, 'Where are we today?' This is a picture that was computed. We computed this last week. This is using the BIS data on total growth of credit. Again, I'm doing this exactly the same way that we do in the paper, which is looking at credit to GDP and then looking at the three-year change. I'm using here the most recent data from the BIS, which goes through the end of the first quarter of this year, and looking at the three-year change. I am not pairing this with asset price growth here. But I think it's worth just starting with credit growth for a second. Now, what do you see here? Actually outside of Asia - so in the US, Canada and in most of Europe - credit growth over the past three years has been zero or negative. It's been very modest. Not shocking, right? That has been one of the goals of central bank tightening to bring down inflation. So, if anything I think when I look at what the data are telling us today it's that we actually have got over the hump in terms of, if you look at where we were two or three years ago, two or three years ago we were what I would've called the yellow zone, which is to say we were massive booming credit markets but not over the top like 2005, 2006. We were just booming credit markets, and of course equity markets two or three years ago, booming enormously. But we actually managed to pass through this without a set of crises over the next - over the most recent period. It's hard to see here but in the very right of the picture you can see that Hong Kong, China, Thailand have experienced massive credit growth over this period. If you look in those countries, the warning signs are a little bit more concerning. To conclude, how predictable are crises? Substantially - and substantially more than we thought. The key thing to remember here is this confluence of credit growth and asset price growth. It's a marker of a credit boom and is it sufficiently predictable to warrant early action by policymakers? We certainly think so. We've tried to make an argument that it is. I will stop there. I'm delighted to take any questions, comments, feedback from all of you. Thank you. Shall I sit down?

Well, that was a surprisingly optimistic start to the day. I love that. Just to let you know what's going to happen, so we've already had - well done to the audience in London - we've got four questions on the slide already, but please feel free to keep them coming in. I'll take the first question here and then we're going to go to Milan next and then Frankfurt so you've been warned. I'm going to put together two questions here that relate to the three-year timeframe. It relates to whether the remaining terms of governments typically are less than three years. Does that mean that surreptitiously they might fail to act? Following on from that, is there a price to be paid from preventing crises?

Two great questions. I think of the key government actors, probably the central bank, and the terms of the central bank are a little bit longer. We hope that they have longer horizons. That said, if you look at the debate between central banks over the past decade post-GFC, I would say there's mixed willingness to take on the avoid-a-crisis mandate. Jeremy Stein, who was a governor at the Fed, has been the biggest - I think - advocate of doing so. His argument is, even if the only thing you care about is long-term GDP growth, we know that crises cause a cumulative massive decline in GDP growth over the next decade. Probably even if all we're doing is using a crisis as a stand-in predictor of the growth part of the central bank mandate, we should be doing stuff. The thing is, it feels a little bit more indirect for central banks. Now, if you look at what - the BIS has been advocating for this for years. The Fed has not really actively taken it on, interestingly. So, the Fed of all the central banks in the world has been the most, I would say, stand-offish on this particular set of issues. So, yes, the costs of a crisis are enormous.

So, worth preventing?

Worth preventing. I didn't go through these numbers because I was running out of time here, but you can do this exercise. I think it's a fantastic exercise, is plot the path of GDP over time. Have time zero be the moment of a crisis and so you have this line of - let's call it log GDP - that's just going up in a straight line. You've probably seen a picture like this, right?

Yes.

Then the crisis happens and it drops. It never really recovers back to where it was. So, you can do this exercise of, it's like the calculus exercise when we were at the end of high school. What is the area under the curve? How much GDP growth have you missed? I think you can convince yourself that the numbers cumulate to something as big as 100 per cent of GDP over a ten-year horizon. Now, I'm not saying that every crisis is that big, right? That would be way too large of a number, but if you look at the averages they're enormous in history. We tend to forget about them and we don't feel that. We don't feel the GDP that we never had, but it's a real cost for society. It's a massive cost in well-being and longevity and so it's a big deal.

Good that we had an optimistic start to say that we're over the hump, as you said.

I do think we're over the hump. I think that if you looked at the data a few years ago, you would've been concerned. It wasn't code red. It wasn't a three-alarm fire but it was a moment where I think we would've had some concerns. Certainly we had our scare in the regional banking crisis in the United States, and of course Credit Suisse. But I think we're over the hump in the sense that we had both a period of sustained credit growth. We had also asked for price growth but it wasn't that hot. It wasn't again over the top. We have had a shock, which is a massive raise in rates in many developed economies. As we know, the banking sector has survived it for now. Yes, I think it's optimistic.

Excellent. Well, we've got lots more questions here in London, so thank you. Keeping those coming in, but I'd like to now go to Milan. Maria, do we have any questions for Robin, in Milan?

Yes, hello Michael, hello Robin. Thank you for the talk.

Hello.

The question we get from - can you hear it?

Yes.

Yes, the question - actually there, you can see me now. So, the question we have from Milan is, I think it kind of echoes what Robin was saying about the Fed being a little bit stand-offish and maybe other central banks less so. The question is, is it fair to say that Europe is less prone to financial crisis, particularly given what European Central Bank has done in terms of regulation of European banks?

I'm not sure I would jump - I wouldn't jump to that conclusion. Capital ratios are lower - tend to be lower - in Europe than in the US. I think the other thing that we have in Europe is that there are costs of the coordination between the different central banks. Having a common currency increases those costs and makes it… This is kind of by analogy: in the United States, if you're having a boom in California and you're having a bust in Kansas, you've still got one Fed. How do you think about that? You have that issue enormously in Europe and so that reduces the power of the main policy instrument, which is the Fed's short rate or the ECB's ability to work on the short rate. I'm not sure that I would agree with that. Certainly in history we've had plenty of crises in Europe. I would say there's more institutional awareness of these issues in Europe at the moment. The BIS, of course, is based in Switzerland and has a bigger voice. But I wouldn't say, if I had to make a bet, that the probability was any lower.

Thank you, Robin. Tim, I guess you might've noticed in Frankfurt that the one crisis that your model didn't predict was the German one. But any questions from Frankfurt, Tim?

Yes, we do. We have a couple of questions on calculation methodology first. Then I'll get to some others after that. Specifically some of the inputs to the model. You used a three-year change in credit to GDP. There was a curiosity whether you looked at the absolute levels of credit to GDP as well as an input and how that went.

We did not - and largely we were going off of prior research that has used the change in credit as the main indicator. I would say this is a contentious issue between the co-authors on this paper, because how do you measure this? It does matter, so for example one of the things that you may have observed when I was putting up my numbers was that I'm looking at the change in credit to GDP. Well, I could also just look at the change in credit. It turns out if you look at credit to GDP it works a little bit better because you're really capturing credit growing faster than the overall economy, which is our goal of what we want to capture. But of course, if you were to use just credit growth by itself it would also be a useful indicator, just not quite as good. Now, the question of credit levels is a good one. Now, if you look over history, you get what people call credit deepening, which is to say that over time - over, say, ten, twenty, thirty-year horizons - as financial markets are developing you just have a straight line of growth of credit in the economy. That in itself means that the level of credit isn't that useful for forecasting crises. It's really the acceleration of credit relative to the overall economy. Yes, credit markets are a good thing. We love credit markets, right? This is very important for the economy, it's important for growth. The popular measure that I've seen on Wall Street many times is thinking about how much - what's that incremental role of credit relative to GDP? How much incremental GDP is it generating? If you have too much credit growth relative to the growth it's generating, that's really the concept that we're after with our measure. By the way, I would say for those of you who are interested in this, you can literally go to the BIS. The way they report, often, their data is credit as a percentage of GDP to start. So, you can pull it as a percentage of GDP. You can pull it in raw terms and dollars so you can get it in all of these different forms. I think recognising that they tell you slightly different things.

Great. Well, let's take another couple from London because I've got a few questions here. Then we'll throw it back to both Frankfurt and Milan in just a sec. There are a couple of questions here that relate to, one that's specific to emerging markets, but they both get to the same question, which is: what role do you think that public sector debt dynamics could play in triggering a crisis?

Most of the time, we think of - Michael, it's a good question. Most of the time we think of the public sector debt issues I think in two particular ways. One is as a response to a crisis. So, for example if you look in Ireland in 2008, 2009 the public sector takes on a lot of the debt when you have a credit contraction from the private sector. In that sense, long term does that create a fiscal issue? Yes, potentially, but that is less… In that case, you're seeing public sector debt escalation at the time of a crisis. But you've got to remember what the causal chain is, right, which is it's the public sector is responding to negative developments in the private sector. Now, I think you might be referring to what's going on, or the questioner is referring to the question. A big issue that's being debated around the world, but particularly in the US right now, is we're running massive deficits as a percentage of GDP, with it looks like no hope in sight of a decline of either the primary deficit or the total deficit. What is the impact of that in terms of forecasting a crisis? If you go back in time - this is some of the work from Reinhart and Rogoff - there's some role of public sector debt in terms of forecasting a crisis. This has been a big theme in the emerging markets, as you probably know. In developed markets, less so, but that doesn't mean we're not going to get there. Most of the economies that we looked at in our dataset are what you would call developed markets today. In fact, only I think ten per cent of the country years in our dataset would qualify as emerging. But I think if you were to do a broader emerging markets analysis, the fiscal entity - the fiscal position - would be significantly more important.

Yes, and that was exactly the question that we got really. Maria, any more questions in Milan?

I have a couple of questions that are quite similar to each other so I'll try to summarise. They are around contagion risk. So, how important it is and how cross-country contagion or within-country contagion, possibly related to SVB Bank collapse earlier in the year when, as far as I understand it, your model wouldn't predict it to be a crisis as credit was coming down. But one bank was falling, another bank was falling so were you worried about contagion and how big an impact it has on overall thinking about crisis?

Yes, so I don't think my work - the work that I presented today - speaks to that issue but I can offer my own opinion on that issue, especially today. I think we have learned some things from the SVB crisis. I think we have learned that the nature of banking crises today is very different than the banking crises that we have experienced in the past. From a personal seat I would say I have a family member who was running a company, and their company had all of their reserves at SVB at the time. So, we had a front seat on what was going on. The way I'd describe it is that that was a crisis that happened over Slack, because it was very fast in terms of just everybody's awareness of the position of the Bank. As you know, it happened over the course of two or three days. If you look over history, bank crises are very different. I don't know if folks here in the audience know this, but when banks are resolved in the United States, it typically happens over a weekend, which is that they are closed on Friday. There's a lot of work that happens on the Saturday and Sunday. Then the goal is to reopen on Monday as a new bank that has been ideally assumed by another bank. Now, we're lucky with SVB that actually - and again this is a little while ago so I'm not sure if you remember - but it was Wednesday, Thursday, then Friday when things were happening. A lot of tension on Saturday and Sunday, and it ultimately resolved on the Monday. Actually the timing was quite lucky, if you think about it, because it wasn't that things had started on the Monday or Tuesday. I think that these things have really accelerated where we are, and that the next set of crises at the bank level will be faster. Now, you could take a step back and say, 'Well, of course people don't want that so what is the implication for policymakers?' Well, policymakers are probably going to be a bit more aggressive. Well, does that mean that we're in a world where the policy is essentially providing depositor insurance to just about everyone? That's a risk and an issue. Maybe that's going a little beyond what we're talking about today but it's a big set of questions

Tim, so we've got four… Well, you can see the clock as well, Tim, can't you?

Yes, it's the very end.

Yes, any more questions from Frankfurt?

We do. I suppose the final question is again about the methodology and the work in the paper, and thinking about the dependent variables that you're looking at, and specifically whether you looked at any capital market outcomes, say, equity returns over the following one or three years, and what explanatory power the model might have to explain that for, I guess, local currency equity market returns.

Great question. We didn't do that. The question was about whether this is useful for forecasting stock market returns. We didn't do that in this paper, but in fact remember that I spoke about this dataset we use for crises, BVX as I call it, so Baron, Verner and Xiong? They have a paper that actually tries to forecast bank stock returns generally and using crisis-type variables. My strongest guess - although I have not done this analysis myself - my strongest guess is that you could use our variable to forecast financial sector returns for these markets, because financial sector is closely linked with what happens in a crisis. There is a separate question of whether it goes beyond and forecasts the overall market during this time. My recollection from the Baron, Verner and Xiong results is that it does. Again, that wasn't our central goal in this paper. We really were writing it largely from a policy perspective, trying to understand. But my best guess is that there will be some market-level predictability. Now, I should say this would be the kind of thing that you would incorporate along with many other things, because crises are still fairly rare events. So, you would want to use this in conjunction with a bunch of other things.

We've got two minutes. Now, this is a rather big question but you've got two minutes to answer it. Do you think that President Xi in China was right to have clamped down on the real estate sector from a credit point of view?

When things are going well and you fear that things will blow up, it is hard to figure out what to do. I have had this debate for a couple of years with Larry Summers about bubbles in general. He is very sceptical of, I would say, the Fed or central banks acting on bubbles because he'll claim… It's a caricature of his views, but it will just say - he'll say, 'The models are sometimes wrong, they're sometimes right. You've got to think that the model is right and you've got to have faith that the policymaker is going to figure it out in time. They're probably going to be asleep a little bit so it'll take them a while. By the time they finally get around to doing the right thing, the thing is exploding anyway and they are clamping down on something that's exploding anyway and they're going to make it worse.' So, I think that's a central question when you're looking at these things. The issue in China - I'm not a China expert - but it has been in a sustained aggressive credit environment for such a long period of time. We could've had the exact same set of discussions about China five years ago - even ten years ago - with massive escalation of housing prices, of equity markets and combined credit. So, I think the issue there is that tension between things that were going to end badly anyway, and are you providing enough cushioning on the runway versus are you able to slow down something so that it doesn't continue getting worse? It's a hard question to answer but I think those are the central trade-offs.

But you've done it brilliantly - and actually you did it bang on time.

Bang on time!

Thank you, Michael.

Thank you so much.