Transcript of the talk, from a Whisper transcription of the audio. Auto-generated: speaker names in particular are unreliable. ======================================================================== # The Very Long-Run Costs of Financial Crises Authors: Karsten Müller, Chenzi Xu, Mohamed Lehbib Discussant: Sanjay R. Singh Video: https://www.youtube.com/watch?v=wEt4iKiNNZQ&t=20775s ## Talk (05:46:15 – 06:30:00) [05:46:15] Okay, so we're going to continue then. [05:46:22] Next session is preparing for the very long cost of financial crisis. [05:46:28] Thank you guys for the program. This is where with my own partner, [05:46:38] Carstin Mueller and a fantastic graduate seat of all. [05:46:42] And it is about financial. [05:46:45] Which is a picture that I've been overall familiar with. So throughout the state where we're going to use financial crises and your changes with banking crises. [05:46:55] Banking crises have been occurring throughout this year. This happens to be my favorite one in the UK in 1866. [05:47:04] But we have more recent examples in the 20th century. And of course, the 21st century. [05:47:11] These events are rare. [05:47:14] But they have been consistently occurring throughout this year. [05:47:19] 2008 recall of the global financial crisis or a good reason. This was the peak in terms of the number of countries that are simultaneously experienced a large banking sector. [05:47:32] If we turn to a lot of data, just looking at for instance, the United States and the United Kingdom during that particular crisis, we see that GDP real GDP did not return to trend. [05:47:46] So here we're just taking a very reduced form of approach and we draw on the timeline in the preceding 15 years. [05:47:53] So it doesn't really matter how we do this. The pattern is very clear. The world that we experienced after the crisis was not sufficient to recover the level of GDP that we would have experienced had we followed this in trend. [05:48:09] There are some historical cases that are actually even more extreme, which is that GDP for a long period space below its previous level. So the world is not insufficient to recover in the post 20 years, the level of real GDP. [05:48:29] Alright, so in this paper, we're really going to focus on each of our counter-factual output gap relative to trend. But I wanted to show you these relative to the levels just to satisfy that in fact, you know, this idea of moving from a worry is not always the case. [05:48:48] Our question is pretty straightforward. It's to ask whether relative losses after crises are a transitory requirement. Another way to put this is how fast is the recovery after one of these large dislocations. [05:49:03] The transitory view, for instance, modeled by Freeman and the fucking model is that after people's sessions there are abnormally large recoveries such that you do actually get back to trend. [05:49:15] And a very striking and fearful paper sort of in this area is the limestine, which shows that after the bombing of Japan, you actually get these remarkable reconversion back to what it would have been. [05:49:30] But then there's other words that argue that these locations are much more long run. Sarah Saxena is the nice example of that. My own John Martin paper had very long run outcomes. [05:49:43] And so, you know, the prior that I ran into there where there's quite a bit of skepticism, maybe really want to write this paper, which is to look at the systematic. [05:49:53] And this is a question that we think is very important for policy because larger output losses justify larger. [05:50:01] So that's not so even when you look at this part of the we're saying you but so that's assuming that's that was optimal or better than the one when you can. [05:50:13] Yes, what do you think the case for instance of the Korean crisis. [05:50:18] And so it stops in the station where you say those guys are going actually in efficiency fast. We're over a community capital effects about that time whether we're doing that that is not decided. [05:50:30] So you're saying that your cases where the group like the traffic, in fact, is not a nice long website. [05:50:38] Yeah, some patients stand. [05:50:42] I'm saying that's not the way for with the case that that's also for example. [05:50:48] So, I agree to know that when there are large financial accumulation and so that can have actually cause a lot of on on bus. [05:51:01] And so that is a blue and I'm really thinking a stand on the way we think about this sort of conceptually is that there is some sort of long term both that country would have been and we just say that longer and both. [05:51:16] Yeah, probably. [05:51:18] We allow for that long term both that to actually be estimated using the white tool. [05:51:27] I'm going to be that exactly intervention is not going to slow it. [05:51:33] Yeah, I mean, so we don't do anything with welfare sound very generously shared with us is his thoughts and paper and another way to think about this is what would be, you know, sort of the, what are the sort of welfare costs of these ones. [05:51:51] All right. So, there's very little on the earth on this because there's a significant native challenge, which is that you actually need very long time series for a large set of countries and the existing evidence, even though the organization tend to focus on short [05:52:09] horizon so there are some people that go out to that 10 years but the median of what most of the results that is around five years or so. [05:52:18] What we find after looking at the this path of about half a century is that well first of all we have to put together a variety of data set in order to look at this. [05:52:32] So we put together basically as much data as we absolutely can. So it has over 200 countries, we put together 46 macro variables, and we do this in a way that allows us to consist of a look at all these articles. [05:52:50] So, we find first that there are larger real GDP and output gaps than all the other estimates the trough is at about 70%. This results comes out of the fact that trough losses occur several decades after the initial onset. [05:53:09] So we find estimates that are very similar to the existing literature when we look at the same horizons as the new literature, but it just so happens that we probably a person much earlier than group has been empirically shown. [05:53:25] Okay. And then in the last part of the paper, we have some evidence on mechanisms. This comes from the kinds and physical and human capital investments because in technology adoption, as well as a it to public finances, which was moving [05:53:47] forward in the political for government to be pushing some of these costs. Alright, so the bottom line of this paper is that the costs of these money crisis are larger because they're more long lasting than as previously been shown. [05:54:06] Alright, so this is a very short report paper. There's a component where I'm going to show you the data that we needed to put together in order to estimate these long run events. [05:54:17] And then I'm going to show you our baseline estimation strategy and how we get our baseline output of that, as well as a huge dimensionality of robustness checks because at the end of the day, this is a paper that primarily has one result so it would be good if you guys believe that would result. [05:54:37] And then finally I'll turn to the magnets, etc. And so what why we would see such long run. [05:54:46] There's a bunch of people that have shown that long-run forms like the financial crisis, especially looking at micro-reason, so that the currency labor. [05:54:58] So we kind of know that the, you know, the system does a fact that you're very much in the process of. [05:55:08] So, for you guys to go into the first time, is it for like a beautiful country. [05:55:14] Well, okay, so I love that you think that everybody's prior it is that these effects are very persistent because at least one I was, you know, working on my first paper. [05:55:26] And I think there's a lot of pushback on this idea that lots of people persistent and you know, evidence in 2008 is not actually that long. [05:55:37] But we can. [05:55:39] So, we're going to go into the long run. [05:55:42] I will show you how our estimates and others and those also are, you know, holding country estimates as well. [05:55:51] But the point is really, let's look at them all. [05:55:56] Okay, so the data we put together for this paper, we studied the financial crisis, we take that from literature we did not have a new crisis, a new chronology that's a very crowded area. [05:56:12] We use the standard definitions of a high crisis where there's why so many times the distress and sometimes government intervention. And so we take a primary about 11, well is the Georgia show our table for Baron, Baron, Sean, and we essentially take a union of the crisis [05:56:33] that they have identified. We choose the dates that earlier in the case of conflicts, and we have an extensive discussion of how we sort of arbitrate those. [05:56:49] It doesn't really matter what an option is. [05:56:54] We also, we also have data on that and we include those as controls in some specific. [05:57:02] So, turning back to this picture where you know by Asian crisis do occur regularly but are rare events. They ordered to study that after the long run and we do we need a lot more data that is typically available. [05:57:16] And the existing sources I would argue all short for many many reasons. So, international organizations put together fantastic data like the I have Leo, for instance, those tend to start in the second major globalization by 1960s 1970s. [05:57:31] There are really nice after going for compilations for instance the word that we're sure our compiler method, barrel or so well but those are focused on specific countries that subset of years as a set of variables, and actually just when we were starting out and trying to put together more of a [05:57:51] data set that was large enough to look at this we found that there were actually widespread errors in terms of, you know, wrong units missing values go to the zero statistical breaks just issues that you know might not realize and to be really try to dig into it and [05:58:09] to be sure that there's some. And so, as a consequence, we could rely on a single resource which wasn't going to be sufficient for our purposes, or we could have made these various decisions about how to sit together these multiple sources. [05:58:26] And so, on Friday, it was fully the Carson was saying, let's do this in a way that is completely transparent open source record, record the wall and continuously updated so that you know the work that we do here which was going to be significant actually [05:58:46] So what we do is we combine 35 sources that are continuously updated these are sources like the IMF the World Bank, where they're sending out orderly or annual updates. [05:59:00] We combine that with what we call in the very two historical sources these are ones that generally after myself with the other four particular papers and the most out there is out there they're not updating it again. [05:59:11] And we harmonize everything for these units, we combine everything and we do that process of ratio splitting, and that deals with large jumps that from just for naively in the data together. [05:59:27] And we have an extensive process for shipping for hours and not only a review in office. So at the end of the day, our coverage starts in a very first year of country year automation in 1086 but really, you know, it's a data set that is about as confident as it can get for the 19th century onward. [05:59:50] Just to give you an example. So really the name of the game was to be as transparent as possible about exactly what we're doing. And for every country and variable we have a full time series where we show you exactly where we're getting all of the underlying data and we show you our data points which are in this lightweight lies where we take [06:00:20] changes in the underlying dominant source and then each of these points is fully documented and all of this together, the place puts together is this open source. [06:00:33] So if you don't like how we've done it and you know you want to substitute it in a different way is that you can totally do that. That's up to you. [06:00:43] So here for instance we have 15 source series that was combined to one we gracious slice of to the OECD and how we outlook in 2018 and that's what gives us the smooth series going forward and back with even as new data is added. [06:01:03] Actually, this one, so there's a series and you're switching it together and putting the point, but exactly ratios. [06:01:11] We made the levels the same and then you're sure the entire data set on one so that they match up to you guys. [06:01:21] Exactly. And so that's something where you have to decide where you want to anchor. [06:01:26] Yeah, so you anchor. [06:01:30] So we anchor it to something like tensioners riders today or like people in private in the rain as many future updates get better because that and prior stuff gets into that. [06:01:41] Yeah, yeah. [06:01:46] sources. [06:01:49] We do not say, for instance, that the I have as it actually take me to that. [06:02:04] We just take the data from where I saw. [06:02:15] Yeah, we actually do have quarterly data for a lot of these articles, and that is in a upcoming release. Yeah, but we started with that. [06:02:32] I think it's very, very good. [06:02:35] Yeah, I have a hard day, but the country's and we are the real forecast that thing this is question is the economies of the various countries in the end of the guys building and they put in the, you know, that thing is very good quality control is the numbers. [06:02:51] And so we have to make a judgment about which data source we prioritize over others. But in the recent period where there's a lot of different sources and there's a lot of the briefing actually. [06:03:06] Yeah, yeah. [06:03:09] Yeah, yeah. [06:03:11] People are going to get the stuff from the I face and they can just ask you. [06:03:16] That's true. So at the end of the day, there's no true underlining source. We're not all we're also just going to be following on a train that there's no true. [06:03:26] You never been a lot of three can I just go to the man. [06:03:29] What's the case. [06:03:30] Yeah, yeah, we have that. [06:03:35] We haven't done that. [06:03:38] We do have a very digitized many store resources for the first time. And this is something that's continuously happening so that as we work with each update, which is quarterly as a we update the data. [06:03:52] We are adding new data so it is constantly being more and more. [06:03:59] Okay, so the only place that you're not working on the name right. [06:04:07] That's right. And what is that in the case but so that's really fairly. [06:04:12] Exactly. Yeah. [06:04:15] Okay, so at the end of the day, we have quite a large sample and for our particular variables of interest. [06:04:24] They, it is by large, but generally speaking, it's a, are, you know, our. [06:04:31] And doing this was to make it as easy as possible for researchers to take this data and apply to whatever conflicts they might be setting. [06:04:42] All right. [06:04:43] And this is just advertisement. [06:04:45] Basically, we try to make it as easy as possible to actually use these data. So there's a website. It gives you all the documentation. You have these nice little figures. [06:04:54] There are also native languages in your coding language of choice and finally, it actually also leads to AI so that you're under graduate. [06:05:02] You can keep this. [06:05:05] Okay, let me turn to the meat of the paper. [06:05:08] So we are interested in estimating very long run best. And we do so with panel local protections taken from also Georgia's terminal work of us. [06:05:20] The left hand side is the cumulative change in real GDP from your prior to the crisis to some horizon H beyond it. [06:05:31] So this is a long difference. And we use all differences in our baseline because they're less prone to bias than LPs and levels. [06:05:41] We have a lesson that shows what happens when we pull the LPs and levels. [06:05:46] Our right hand side is the same crisis in the mirror that we have taken from the literature that equals one in the first year of any crisis. [06:05:55] We only part a country year observation as having no bank crisis. [06:06:02] If we know for sure that these one of the chronologies examines that country here. So we're not filling in missing data with zeros. [06:06:15] The back in person zero one. [06:06:18] Yeah, but then you guys in zero one, we do not use this. We know he's the bar in front of strong, you know, I believe the crime. [06:06:26] No one of them. They may be different. [06:06:29] They can say, SVB banking credit which is a part of the bank and better than believe my death. [06:06:35] Totally. [06:06:38] Yeah. So the hero and I think that's that was one of the big contributions of the bar in front of show me for which was to say, in fact, it is important to quantify the sense of a crisis here. [06:06:52] The vast majority of these chronologies are just zero one that's what we do for now. The one place where we think about the intensity of crises is where we can think about pricing that are sort of in these large world shop. [06:07:07] So the great question is large, local shot, the data is large, local shot. Those may be quality for very different from other ones, but that's, again, not like at the country, my year. [06:07:37] Okay. [06:07:48] So our goal to the interest are these days estimated at each horizon. So these are really estimated regressions. So it has to be humor to the output difference is the gap initial on controls. So for example, the state of 10, sorry, based on 10. [06:08:06] So at horizon 10 or point of set of these we have an out the gap of 7% of GDP. [06:08:13] What the reason we really like these little projections is that this is a completely reduced form of approach where we let the data to see a sense of a person. [06:08:24] We have 100 picks the best to do our long run. [06:08:29] Well, we have a first difference is already a sort of long run. [06:08:32] But this absorbs differences and trend growth across country. [06:08:39] And a key thing that we do is we include 20 in our baseline 25 last of output quote. And so that's is missed this entangle the post crisis trend from other things that are have like the trading would have been upon from the [06:09:00] So one hypothesis. [06:09:04] The country developed to go on with fast growth and when the possible like in Korea, they have a financial crisis and then they mature and then gross about that. [06:09:14] So, but that's not just because of the prices just the process of development that that is, it's the crisis and so once you get in the development to have a growth less fast. [06:09:26] And then you maybe put like a graphic trending growth to allow for longer than that to change your in your robust and this type of thinking. [06:09:35] Yeah, that's a great question. We haven't done anything like that, but we can look more at first on set or on set of financial like the essentially versions of cutting the data for her to make it up to what where it can treat as at the time. [06:09:53] Yeah, actually, my class and he's been saying that he said that he likes and I'm doing much for the videos. Right. [06:10:01] The kids like to control the design of the mobile company to say so I think my mistaking or accepted that the split prices and that should know that the person of purchase 50 liners enough, right. Yeah. [06:10:13] So, there are. Yes, there's a little world for guys long as you have enough of the last thing that that's sort of like a half shirt or these different growth dynamics, then we kind of absorb those. [06:10:27] But that's a different thinking about the head or she may be of, you know, if these prices will very different. [06:10:35] If it's the first period or not. [06:10:39] Like, yes. [06:10:44] Yeah, and I mean, that was required for reviews and data for a few of our classes right so like, [06:10:51] Yeah, but then [06:10:54] you just do some more of the freedom. [06:10:58] But then you're going to have to assume some kind of like these individuals are poorly. [06:11:04] Let's talk about the literature. A lot of these prices are all one pool of that. [06:11:11] So, for us, actually, if we are worried that there is some serial relation within years, or, you know, within countries. [06:11:22] That's why we need to stress will bring throughout in our baseline, which is meant to be robust to serial and cross country correlations. Now there's some work by [06:11:34] that said [06:11:36] Oh, sorry. [06:11:39] Okay. [06:11:44] With last year, which is what we do robust standard hours and even cross section is going to supply so that's going to correspond to close to my year and so we also show you your last night. [06:12:00] And pretty much throughout your last year, you see smaller confidence and our whole so we think the D. A. is more conservative confidence and our whole which is what you're going to use in our discussion. [06:12:14] Okay. [06:12:15] All right, so because long run estimates are going to be noisy, we did need this much largely data set to address different elements of bias we have we [06:12:27] have to reverse what the L. T. [06:12:29] What is your most thought is we just use one difference in this country, this affects large lives. [06:12:35] And there's some recent work showing this is getting really to the, the police and things but there's some recent work showing that there might be some legal bias when it comes to L. [06:12:46] And so we do a version that has a jackpot and show the robustness to that version of this meeting and the throughout the school. [06:12:59] Okay, so here's our main result. [06:13:02] In the first five years, we see losses that look about 7% of GDP. And that is very close to what the rest of literature has shown where it comes much more traumatic is that these losses basically continue to trend down on a super significant way, but they do continue to accumulate [06:13:26] in sugar three to four that get horizon before finally starting to sort of revert in the much later three of our sample. [06:13:35] And so these that are there since red ones under this whole credit the orange ones are the near clustered. [06:13:45] In the full panel, where we use all of our data, the number of observations that are identifying these different arising latest is changing as the data sort of becomes more scarce as we move out. [06:14:03] And so in the background of these days is also a composition on change of what's going to show up with Christy's which countries, etc. [06:14:13] And so to let a little bit of clarity on that. We also throughout the paper, we're before to the results for what we call a balance panel, which is where we follow the exact set of crisis for 50 years and so that requires that we see the full set of outcomes for the 50 years [06:14:33] exposed. This also means that by that mission we're not including the most recent crises, because those would only take us out of 10 years 20 years, etc. [06:14:44] When we do that, we have a deeper pattern where, you know, the losses continue to accumulate a ratio dropping at your 30, before starting to rebound. [06:14:57] And this is where we get our 17% number. Okay. [06:15:01] So, one way to read the difference between this balance plan, so I heard just one second, and the whole panel is that the more recent crises, especially in the shorter term have less severe effects than we've seen in sort of a broader panel over. [06:15:19] Sorry. [06:15:22] I'm not looking at data. [06:15:25] They don't get a crisis and I'm clear whether they're not crisis. [06:15:28] Yeah. [06:15:30] So, okay, so in the baseline, we don't do anything with it. So, each new crisis just registers as a new onset and then the path of GDP followed us. [06:15:42] So, there is something we do here where we where we can control for subsequent crises, we can also control for proximity to the most recent crisis, and we can control for the lack of the end of the month variable. [06:16:01] And actually, when we do that. [06:16:05] I get through with it. [06:16:08] It generally, it generally looks the same, but it's actually a little bit deeper. [06:16:14] Yes. [06:16:16] But what you're doing by crisis, the crisis may have a drop. [06:16:53] I think that's another way of asking thinking about the energy of the crisis intensity. [06:17:08] The energy of the, you know, how long the actual year of this location itself last. [06:17:15] So, we have not used that data, but that's something that we should sort of think. [06:17:23] I do think though that you know, no matter what everything the crisis in like one year or five years it's hard to sort of saying, yeah, that's, but it would still be an issue. [06:17:35] Um, do you think it's been a part of the like, like minus horizon. [06:17:43] What would that mean? [06:17:47] We. [06:17:49] Okay, yes, we did look at this. And without last, you do see a somewhat of a boom, which is what he really says, once you start to realize it. [06:18:04] I actually think he doesn't I guess very good. [06:18:07] You can rate the event, but by the distance, what we're most about them to be really say, and then it becomes an exposure man and then it's much more. [06:18:18] I think it's a little bit of a change in the organization. [06:18:24] But it's a, you know, the least sense means that every country has a different. [06:18:40] So, we do a lot of things here that you click on this month and look at some more pictures if you want, but we also summarize basically this massive amount of robustness that would do the I'm going to show you an incredibly dense people so we're going to do this bit by bit. [06:19:02] The way to read this table is these three columns are for horizon, 10 estimates, the beta and the T staff using your cluster versus fiscal prey. [06:19:13] And the little stars are the components, sorry, the, the, the significance for the success. So you all know. [06:19:24] Okay, so this is at horizon 30 years and this is kind of part of the main point of the paper, which is that things keep getting worse before they get better. And then finally, the last three columns we show you the truck. [06:19:39] So what is the truck and then what horizon that occurs. [06:19:45] So now we re estimate everything where we can include controls for contemporary is dislocations wars, the grants and sovereign debt crisis. We can control for macroeconomic outcomes. [06:20:02] This is the proximity to other crises or the entire path of crisis history. [06:20:10] We can include a control for the world should be gross so this idea that you know there are just some really bad years and then that's the one side of the crisis and that's something that we'd be pushing up instead of the real and happy crisis itself. [06:20:25] So I think the big success, I have got some message, you want to get into it that makes the best other sort of saddened controls like excluding the world wars and the great depression. [06:20:37] Using these individual chronologies and additional lives levels. [06:20:43] So this is the version where we look at everything in real cheaper capital, etc. Right. So, I can show you as a little else like this where we look at what you need for a comma, or with the balance panel. [06:20:56] And we also have a placebo where we randomized the crisis events. Yes, may ever. [06:21:04] So this is meant to give you a sense of the robustness and also a sense of where you know the results are a little bit more heterogeneous. [06:21:15] We don't have those various system and we don't have that in our. [06:21:25] Oh, yes, okay, so. [06:21:31] As we saw in the chronology of that crisis, there are some years where there are basically global financial processes. [06:21:40] 2000 years one, 1931 is another, 1907. There are just periods where the death of a financial crisis just is, I think, much bigger than other periods. [06:21:55] And a year face effect is going to absorb that heterogeneity. [06:22:01] And so it says if you actually do think that this is kind of where I sat that there are some crises that are worse because they are some opinions. [06:22:13] It's happening. So, you know, right. [06:22:15] It's constraints are getting work in the world. [06:22:18] Entirely, then this is basically over controlling. And I think that the fact that it shows up much more intuitive is kind of nice. [06:22:29] That's a great impression. [06:22:34] Yeah, here. Yeah, so actually, yeah, exactly. [06:22:41] Welcome. Okay, so I'm going to show you how our results can be [06:22:47] compared to the literature. This is a fairly broad literature review, not everybody is doing, [06:22:52] you know, VARs or chemical projections. So some of this work, for instance, by my robot, which we [06:22:57] put it on all of the points, is just, you know, they're looking at, they're measuring trough [06:23:01] to peak, you know. So we're trying to distinguish a little bit on the methodology, but the [06:23:07] points here is to show that for a range of estimates, and then the star clue is the precision weighted [06:23:15] mean of the existing estimates, we're very much in the same ball run, you know, 7% of real [06:23:23] sugar key losses in this medium horizon of like five to 10 years. Where things look very different [06:23:31] is when we look at the trough or the the peak output losses, where the big, [06:23:38] where most of these papers don't go out far enough, relative to what we're doing to capture [06:23:43] these much larger loss of evidence. Okay, so if, [06:23:50] If on that, and I guess, for anybody that's, [06:23:53] kind of surprising that, I mean, I guess I can infer from the, since they're all using [06:23:59] different periods that the very stable coefficient forward to to the, [06:24:06] the degree one, I would expect that these such structural parameters that depend on [06:24:12] what kind of stabilization policies are available, what kind of, how the whole [06:24:16] of the bank instance is to the, these things are all changing over time. So our [06:24:21] perspective is very, you know, did you check like some samples, see if there's a [06:24:27] little bit of. So we went down the direction of trying to understand, okay, so one way to [06:24:37] think about mechanisms was to use cross country energy chain, where precisely we were trying [06:24:42] to get a sense of how our dependence or how by the deep countries work, we [06:24:49] didn't get very far with that, and so the measures, especially around historically, were not that [06:24:54] great. So we could do things like looking at emerging markets today as the kind of the [06:24:59] least inviourced is advanced economies, etc. And they're don't see much energy maybe, because [06:25:04] you know, maybe those cuts of the data are really what the underlying thing is. Yeah. [06:25:10] But we would like to be more there, if people are passing out. [06:25:15] All right. So what we do with mechanisms is very similar to what we do with the main [06:25:24] neighbor, which is just to make a completely reduced form of approach and look at the [06:25:31] indicators that we think matter for our countries. And we have really nice data on [06:25:37] the capital investment. We have this really cool data set on technology adoption, [06:25:44] the sort of looks at how the major option and small and then public finances. So we go through [06:25:50] the PPPs and try to use when you're visiting the political economy. [06:25:55] And in fact, something was just funny about that, that we did some, you put our data [06:26:01] with some political economy outcome and it looks like we'll get by major on our way. [06:26:06] Unless we've done a lot of American, yeah. [06:26:18] That's right. And you know, we would actually like so in a sense these, I think, you know, [06:26:23] because this is a very standard growth model, we, if you invest less than the inputs, you would [06:26:29] expect out the past, but the sense of which the public, the government and then you'll be [06:26:33] then how would you, you know, the policy would in a sense is something that is [06:26:40] very also relevant and we'd like to look more. [06:26:46] Okay. So we see that the investment in the blue falls much more significantly than the [06:26:55] assumption does and stays at about 10% until about three decades in. I don't want to reach [06:27:02] too much into these figures, but you know, the recovery of investment is around when you see [06:27:07] the recovery and output. And this is very consistent with prior literature, especially coming out of [06:27:13] 2008 that an investment gap will flow or capital stock has a persistent impact on the output. [06:27:21] Okay. We look at capital generation for workers and technology adoption [06:27:29] and see as you would expect lower labor opportunity in terms of output for our [06:27:40] again. [06:27:42] That we like to go back and buy this company that is components and [06:27:47] we do not do both the country. If that is something people like in the win scene, [06:27:51] we can't share both the county. Yeah. [06:27:55] Oh, yeah. [06:27:59] Sorry. [06:28:02] Like, [06:28:07] just there's one crisis and then 30 or there can be no crisis. [06:28:12] Oh, this is the whole sample. [06:28:14] The whole sample not about this time. And it's just in the scene. [06:28:19] You know, we have these out but whether we have the sound and data. [06:28:23] Be correct. If you like this that long, you will be the thing with this. [06:28:28] Ah, yes. I mean, at the end of the day, these are just kind of like averages over crisis. [06:28:34] Right. So, but we can do more case studies with these individual outcomes like I did with real [06:28:41] G. [06:28:43] Yeah. [06:28:43] No, I did option. [06:28:44] That's where it was like, they want to be one that's [06:28:48] Yeah. [06:28:49] So what it is that makes it. [06:28:51] Yeah. [06:28:51] Oh, sorry. [06:28:52] Right, right. [06:28:53] So, you know, I would not exist, but it's okay. We should work with care. [06:28:58] All right. [06:29:01] There's a really nice thing is that by barrel lead that looks like this historical schooling. [06:29:06] There's a little bit of a contraction and investment in human capital. [06:29:11] And then finally, this is my last slide. [06:29:14] We see that public finances usually admit. [06:29:16] So we haven't done much here, but you do see this decline in revenue, even beyond GDP and deficits [06:29:26] increase. [06:29:27] And so you have large debt throughout us, but then countries generally take out that this over time. [06:29:34] Okay. [06:29:35] So, I don't know. [06:29:38] We built this data set to ask out the system are financial crisis. [06:29:44] We see that they do terribly disrupt output in the drivers of the potential output. [06:29:50] And our takeaway is that the missing estimates generally understate the cost. [06:29:56] Thank you very much. ## Discussion (06:30:00 – 06:45:45) [06:30:03] You've got some Sunday. [06:30:05] Thank you. [06:30:06] Okay. [06:30:29] Okay. [06:30:59] So I think I'm going to write this to the skateboard. [06:31:03] I really enjoyed doing the table. [06:31:05] So very related to some of the work I've done. [06:31:09] But first of all, I want to talk with the [06:31:12] this is kind of the office of the together this large collection. [06:31:16] It's really useful. [06:31:17] And I think for saying, you know, I was very impressed that I could get my [06:31:21] Chloe and then they could talk and then they could do some stuff with it. [06:31:27] So I have one place around the deficit. [06:31:31] It's really cool. [06:31:33] So my discussion brought me on why we care whether what we knew from [06:31:44] previous studies at five, six or 10 year horizons was that there was one with one [06:31:48] 50 year horizon. [06:31:49] If you care about that. [06:31:51] And the second thing is how far this evidence can can take us. [06:31:56] And I think some of the things I'm going to say here seem to drop that cotton did not have [06:32:01] this good for errors, but now they have to put the errors. [06:32:03] So if it's a rest, some of this here. [06:32:05] So largely, I hope it's going to be why we care about long horizon errors. [06:32:10] So there is a regression. [06:32:12] It's running against the local projections for a country I horizon page. [06:32:18] I think the horizon minus one country needs to take some of that and then there's some financial [06:32:22] prices dummy variable and take that basically investing into the cumulative response of the [06:32:29] country. [06:32:30] Do you need me, for example, relative to what the more than the horizon minus one [06:32:36] several eight years after the price. [06:32:39] So what I did is the very George should read to the database. [06:32:44] And this is 17 or now 18 advanced economies from 1870 to 2019. [06:32:52] And they have a label Valencia price is the definition and basically classify a peak as [06:33:00] when you people was high just before this crisis. [06:33:04] So from that, we can work that case out this effect. [06:33:07] And what we see is about 9.5% points lower at your job. [06:33:12] After this price is very large defense event. [06:33:17] So here I'm using four animals were flat country. [06:33:21] This affects and I was moving more involved. [06:33:24] I went and this is where I have done some work on this with Oscar and then Alan [06:33:29] Taylor, where we look at one of the techniques in the same data set. [06:33:33] Finally, persistent effects of this module shops. [06:33:36] The total thing I was looking at is as we just do data set, [06:33:41] federally admired, you learn to share children constructed where they take the [06:33:46] school rates of what part of the town was and they face how they think so. [06:33:50] What's at this was remind me of the previous one. [06:33:55] I have that things are talking about bombings happen with the people sessions, [06:34:00] but it's a good recovery. [06:34:02] Here's some of the recovery, but that's the some persistence that there from [06:34:06] right to shops in the historical data. [06:34:11] Well, I took from this is that there's also the way far from the loopers [06:34:16] calculation of 0.05% cost of business cycles. [06:34:20] If these shops are hitting the economy and these are financial sessions in the [06:34:24] Genesis data set is basically around 3% frequency of country to the financial [06:34:29] prices and that means a lot of the consequences of these events. [06:34:35] So what I'm going to do is first is very simple. [06:34:39] We will look at the calculation of the law of utility that has 1196 [06:34:43] frequency of crisis of 3% annual. [06:34:46] You can translate the calculation of Lucas into frankly, [06:34:52] if the local projection estimate from after price is the horizon estimate [06:34:56] and possibly that into how much consumption you need to give up, [06:35:00] if you don't want to do it, you don't want to do these prices. [06:35:02] So that's the number you can come up with as with copies is what I'm going to talk about. [06:35:08] So here, particularly George, a retailer financial price estimate at year 10. [06:35:14] And then I postulate that the economy kind of level turns back the same trend [06:35:21] from the same level, but year 25, the cost of business cycles on this [06:35:25] recession is 2.8% is one of the ones that are not. [06:35:28] So that's 0.05% versus 2.8% is pretty large. [06:35:34] But what would happen if this was to stay constant and we never had a recovery in the levels? [06:35:41] And this is the number, this is like one place in the solution, [06:35:46] Armstrong is then the utility to Satsena in 2005, [06:35:51] just at Mark's publishing the emerging market site, [06:35:53] there's a friend paper, they were looking at Asian crisis and they said, [06:35:57] founded growth dynamics in the range of recovery countries won't recover [06:36:02] after this deep crisis, especially after financial crisis. [06:36:05] That they were estimating using at the heart for more over time. [06:36:09] Now we can do them directly. [06:36:11] And this is the kind of path they were projecting. [06:36:14] They're single, they just never are recovering levels. [06:36:16] And that's the world we live in. [06:36:17] That's like a pretty large, very consequences of his current. [06:36:21] Now first scenario is well, what it is just a simple recovery from here on, [06:36:27] and the boss comes to me in the middle of the book. [06:36:30] The point being, the mystery is, it matters what these dynamics look like [06:36:35] to trace out the consequences of these of these of these obsessions. [06:36:39] Now here I'm just using the approximate consumption. [06:36:41] So when people have done with ILS at the end, [06:36:44] they found broadly the same pattern for the interruption. [06:36:48] That's my cross-sector infrastructure in these blood glasses. [06:36:54] Now what I want to think is, if I just keep on now standing with GST in post-response [06:37:01] to five years forward, what do I get? [06:37:03] I was hoping I'd get recovery, but I ended up getting, [06:37:06] this system looks like there's a growth in level, [06:37:10] but it's not growth, growth, growth is happening, [06:37:13] so there's no level recovery that's happening. [06:37:16] Well, what's happening is I'm using less and less number of pricing limits, [06:37:20] and my standards are not still there. [06:37:23] It's not, maybe one of the more common processes is also as much, [06:37:27] because you're using less and less of data, [06:37:29] so you estimate them, and then it's going to really be five. [06:37:32] This looks like the deep recession, [06:37:34] changing had about 50% drop in at 25. [06:37:39] Now I can look at the most responses Chenzi has, [06:37:44] which is this is a balanced panel from the people with the conflict-lustrated standard errors. [06:37:52] There's the same drop here, here 25, for 15, 16%, [06:37:56] but then this looks like there's a recovery that we want just from. [06:38:00] The advantage of using this is using much more data, [06:38:03] like this one, 130 crisis events in the dataset, [06:38:07] that's what we're going to be showing. [06:38:12] So if you put this number, the value of the cost comes to be about 5%, [06:38:16] that's what this is, the reduction proxy comes to your 5% of consumption. [06:38:20] Now this number looks basically close to, if we have stopped at a year 10, [06:38:30] you basically, the coming growth rate will be 5%, that's the number. [06:38:34] I mean, this is my simple toy model, but that's close, [06:38:36] you get closer to the number, the value of your stop at a year 10, [06:38:41] and the reason being, the simple model has this counting, [06:38:45] and as you get further and further out, you don't hear as much about [06:38:48] how much loss happens. [06:38:49] So knowing up to year 35 is more or less enough for how much the cost of these crisis are. [06:39:00] Now this is, this one came up a little bit in the discussion, [06:39:04] in the question and answers. [06:39:05] So in their balance crisis cohort, where they are restricting a sample to be before 50 years, [06:39:13] before, so it's like this 1975, the last crisis, then the US, [06:39:16] you get decoder sections there, about 15% drop, [06:39:20] where you could allow these things to mediate, [06:39:24] what I used to be estimated using the whole dataset, [06:39:27] you cannot do for the estimate, which is 25, [06:39:29] 15 data to 2,000, you have 25 years in more prices here, [06:39:34] you get somewhat less data, less data for the sessions, [06:39:38] basically, and that can change the cost of the site of 3-3, [06:39:41] because it can go from 5% to 3.5%, depending on how precisely [06:39:46] you're estimating or how much data you're using to estimate this. [06:39:49] Now something that's already interesting here was, [06:39:52] so this panel is using data crisis under 1975, [06:39:58] this is using crisis under 1999, and the number of crisis that changes in the middle are about [06:40:04] basically 140 crisis that come up, and those are, [06:40:08] a chunk of them are coming from emerging markets, [06:40:10] where there's very short recovery seems in the institution. [06:40:14] So looking at the cost of actually getting to that, [06:40:17] meaning all of these sample splits would be, I think, a useful way, [06:40:22] because I'm admitting to me, this question is super important, [06:40:25] the benefit cost of these numbers is still, [06:40:28] we're now at 0.05%, we're at 3%-4%, but these are very large numbers regardless. [06:40:37] So the second point I want to talk about is, [06:40:41] now, global histories, there are traces of these over long horizons, [06:40:47] but the question that often comes up is, [06:40:50] are these some conventional confidence intervals that we have, [06:40:52] are reliable, when each horizon that we're looking at is large, [06:40:57] that is a sample size that we have, especially when the data is nearly [06:41:01] the real, the real data that we have. [06:41:03] So let me give you a simple toy example to play with. [06:41:07] Suppose we observe the shock, epsilon G, this would be a crisis shock, [06:41:11] this would be a war shock, and this is a universal process and a test degree. [06:41:17] I know that after a little shock, epsilon is a response on the function to be 1. [06:41:28] If I estimate a little bit of predictions here, [06:41:30] now I need to set it here, this true data should be 1, [06:41:35] but in finite samples, this is what my model of talking about, [06:41:40] this estimate can be other than 1, because in finite samples, [06:41:44] the shock can be forwarded by other shocks. [06:41:46] And there can be covariance between the shock and subsequent shocks. [06:41:50] And as it stands up horizon, like after 10, 20, 30, 40 years, [06:41:54] you're now getting 50 years of shocks that can happen in the middle, [06:41:57] and those 50 years of shocks can, in fact, be the initial shock, [06:42:00] and then push it away from 1. [06:42:04] So this is a well-known problem in the literature. [06:42:09] And so a numerator here is basically, [06:42:11] if a original shock happened to be followed by other unfavorable shocks, [06:42:15] that can generate noise in estimation, [06:42:19] and the numerator tells you how much variation you have to find with the coefficient. [06:42:25] And so this H over T problem basically, as we see, [06:42:30] it shrinks this estimation noise here, but as well as the estimation horizon, [06:42:36] and then that can cause, cause, cause, cause, cause, cause, cause problems. [06:42:38] So now in the paper by Jose Montiorega and he can find one model, [06:42:44] we're going to talk about local population in the system for the 18th. [06:42:48] They have this very particular case with the unit foods, [06:42:51] and time period of 30 or 240. [06:42:55] And now if you increase the horizon, basically beyond 15% of the sample, [06:43:02] your coverage ratio of 90% of the normal confidence band, [06:43:06] it will raise pretty heavy. [06:43:07] With some good books, this is the usual common standard influence that we have, [06:43:12] and that that coverage ratio can, can, can be. [06:43:16] So 50 years for 200, [06:43:21] 150 lags, maybe 50 years over 150 years of data, 50 horizon, you had 0.33. [06:43:28] So that coverage ratio is going to be pretty, pretty bad, right? [06:43:30] 25 years over 150s, you're kind of losing 0.15. [06:43:36] So my, my worry with this is how much [06:43:39] reliably we can think about, [06:43:43] because the, the estimates, [06:43:46] stress as, as we try to learn for, for longer. [06:43:50] So I did not, many countries have, [06:43:54] but my worry here, and this is where like my suggestion then was who, who this could pay, [06:43:59] I think so many of, so I think we go back to these hundreds in the data, [06:44:06] most of the countries are still the traditional data countries. [06:44:08] 18 countries for which you have data from 800 onwards. [06:44:12] Really from 1920 onwards is when you start having countries, [06:44:15] 34 countries, 45 countries. [06:44:17] So from 1920 onwards is where you have like an 80 year period [06:44:21] to estimate these impulse responses. [06:44:23] So basically have one or two observations and that can help, [06:44:26] but I think it's useful if the office could [06:44:30] provide some sort of progress, just if you can simulate the data, [06:44:34] simulate like a model based on the sort of structural movement [06:44:40] and show whether the bumpers and are good in coverage. [06:44:45] So that's what I'll stop. [06:44:47] Finally, I think I wanted to say, I find this really, [06:44:53] I think the last slide doesn't want to come. [06:44:57] So the point, I think it's super interesting data set. [06:45:02] I think there is much more value in the data set that currently [06:45:09] currently exploited. [06:45:10] For example, I played around with combining the sorts of what data set, [06:45:15] the four layers of what it said they have to be done in terms of clearness, media freedom, [06:45:18] traditional dependence. [06:45:19] And if you look at this, the demand prices have very persistent effects on judicial freedom. [06:45:25] And we kind of know this about populism, [06:45:27] financial crisis meant to be for biologues or populism, etc. [06:45:31] It's incredibly to me that laws don't have that persistent effects. [06:45:36] Something about how we think about these becoming costs or how they can have much more [06:45:41] persistent effects on society and political economy, they can go beyond the economic effects. ## Q&A (06:45:45 – 07:01:47) [06:45:47] So it's worth saying that in detail. [06:45:49] Thank you. [06:45:50] In percent. [06:45:54] Well, I think that's pretty good. [06:47:03] I think that last slide, but I wouldn't be able to do the full video. [06:47:16] Let me know. [06:47:16] Shall I let you do it? [06:47:18] But first of the present. [06:47:20] And [06:47:43] I don't know why people have trouble. [06:47:45] So we're unable to shoot a lot. [06:47:47] I have no idea. [06:47:48] We should come back here and just click on it. [06:47:51] Yeah, I mean, so [06:47:54] I'm just [06:54:17] so [06:54:35] someone [07:01:06] you can just take that back and watch the video. [07:01:08] Yeah. [07:01:10] Yeah. [07:01:12] Yeah. [07:01:22] Yeah. [07:01:26] Yeah. [07:01:29] Yeah. [07:01:31] All right. [07:01:33] Yeah. [07:01:35] Okay. [07:01:37] So, let's get started. [07:01:42] So people, [07:01:44] That kind of had to be as soon as. [07:01:46] Next session and.