Why Are Recessions Associated With Financial Crises Different?

Transcription

1 Why Are Recessions Associated With Financial Crises Different? Luca Benati University of Bern Abstract We use Bayesian time-varying parameters structural VARs with stochastic volatility to investigate the specific dimensions along which recessions associated with severe financial crises have historically been different from normal recessions, and standard business-cycle fluctuations, in the post-wwii U.S., Japan, Euro area, Sweden, and Finland. We identify four structural s by combining a single long-run restriction, to identify a permanent output as in Blanchard and Quah (989), with three sign restrictions to identify demandand supply-side transitory s. Evidence suggests that severe financial crises have systematically been characterized by a negative impact on potential output dynamics, which in two cases Japan, following the collapse of the asset prices bubble of the second half of the 98s, and Finland, in the early 99s appears to have been nothing short of dramatic. With the single exception of the Euro area during the recent crisis, all financial crises analyzed herein have also been characterized by a specific pattern of demand non-policy s, expansionary during the years leading up to the crisis, and contractionary following its outbreak. On the other hand, no other macroeconomic feature has exhibited any systematically different pattern, during financial crises, compared to standard macroeconomic fluctuations. Our main conclusion is therefore that, once controlling for the size of the s, financial crises appear to have been broadly similar to standard macroeconomic fluctuations along many, but not all, dimensions. Keywords: Financial crises; Bayesian VARs; stochastic volatility; time-varying parameters; structural VARs; long-run restrictions; sign restrictions; monetary policy; monetary regimes. This paper is an outgrowth of my discussion of M. Bordo and J. Haubrich (), Deep Recessions, Fast Recoveries, and Financial Crises: Evidence from the American Record, at the Swiss National Bank s research conference Policy Challenges and Developments in Monetary Economics (Zurich, - September, ). I wish to thank Joseph Haubrich for useful conversations. Usual disclaimers apply. Department of Economics, University of Bern, Schanzeneckstrasse, CH- Bern, Switzerland. luca.benati@vwi.unibe.ch

2 Introduction Reinhart and Rogoff s work on the history of financial crises has established two key facts. First, financial crisis, which, during the years leading up to the Great Recession, had routinely been regarded as a largely out-of-date topic, have been a regular feature of macroeconomic fluctuations for hundreds of years. Second, severe financial crises have typically had dramatic and long-lasting macroeconomic effects, with large and protracted falls in output, and sizeable increases in the unemployment rate. As Reinhart and Rogoff (henceforth, RR) put it, [...] the aftermath of banking crises is associated with profound declines in output and employment. The unemployment rate rises an average of 7 percentage points during the down phase of the cycle, which lasts on average more than four years. Output falls (from peak to through) more than 9 per cent on average, although the duration of the downturn, averaging roughly two years, is considerably shorter than that of unemployment. Although groundbreaking, RR s work, being exclusively based on either the analysis of the raw data, or simple regression techniques, cannot provide answers to important questions concerning the specific dimensionsalong which recessions associated with financial crises have historically been different from normal recessions and, more generally, from standard business-cycle fluctuations. In particular, Have financial crises simply been characterized by larger s, or is it possible to detect changes in the relative importance of different types of s? Once controlling for the size of the s, is it possible to detect differences in the transmission mechanism, compared to standard macroeconomic fluctuations? More generally, is it possible to detect a set of features which have consistently been common to all financial crises, across countries and, possibly, across monetary regimes? The very nature of these questions implies that the methodology to be used in order to successfully tackle them should allow for time-variation in both the size of the s, and the way in which they propagate through the economy, and it should allow for the identification of a set of structural s. See, first and foremost, Reinhart and Rogoff (9). See RR (9, p. ).

3 . This paper: methodology and main results In this paper we use Bayesian time-varying parameters structural VARs with stochastic volatility to investigate the specific dimensions along which recessions associated with severe financial crises have historically been different from normal recessions, and from standard business-cycle fluctuations, in the post-wwii United States, Japan, Euro area, Sweden, and Finland. We identify four structural s by combining a single long-run restriction, to identify a permanent output as in Blanchard and Quah (989), with three sign restrictions to identify demand- and supply-side transitory s. Our main results can be summarized as follows. First, a consistent feature among all the severe financial crises considered herein is a non-negligible negative impact on potential output dynamics. In two cases Japan and Finland in the early 99s the impact appears to have been nothing short of dramatic, with an obvious, significant decrease in Japan s trend output growth following the crisis; and a sizeable permanent output loss, but no significant change in trend output growth, in Finland. As for the recent crisis, its impact appears to have been comparatively mild in the United States and Sweden, and to have instead been quite significant in Japan, with (based on median estimates) a sizeable one-off permanent output loss equal to. per cent of potential GDP. The impact on the Euro area s potential output dynamics, on the other hand, appears to have been intermediate between these extremes, with potential GDP estimated to have temporarily decreased by. per cent from the peak of 8Q to the trough of 9Q, before resuming its upward trend. An important point to stress, however, is that with the single exception of Japan s crisis of the early 99s our evidence does not suggest that, under this respect, financial crises are in any way special, compared to normal recessions. Rather, our results are compatible with the notion that financial crises have been characterized by sizeable negative effects on potential output dynamics simply because, historically, the have been associated with extremely deep recessions, which have ended up blasting away pieces of the economy. Both in the United States and in the Euro area, for example, the recessions associated with the disinflations of the early 98s, which had been significantly less severe than those associated with the recent financial crisis, had indeed exhibited, according to our results, vastly milder, but still statistically significant negative effects on potential output dynamics. To put it differently, with the single exception of the Japanese Although, throughout the entire paper, we talk about the negative impact of financial crises on potential output dynamics, it is important to keep in mind that, strictly speaking, the methodology used herein does not allow us to establish a causal relationship between the crisis and concurrent and/or subsequent developments in the evolution of potential output. Our circumstantial evidence, however with all the financial crisis analyzed herein having systematically been associated with either a temporary or a permanent deterioration in the evolution of potential output is remarkably strong, and it naturally suggests that financial crises have been the underlying cause of such deterioration.

4 financial crisis of the early 99s which impacted on potential output growth, and is therefore in a league of its own what seems to matter is the depth of the recession, rather than the fact that it has been associated with a financial crisis. Second, in most, but not all cases (with the most notable exception being the Euro area during the recent crisis), demand non-policy s which, within the present context, should be expected to capture credit market disturbances displayed a consistent pattern, expansionary in the run-up to the crisis, and contractionary following its outbreak. Third, quite surprisingly, a temporary increase in macroeconomic volatility does not appear to have been a robust common feature of financial crises. For example, for neither Japan, Finland, nor Sweden the financial crises of the early 99s had been associated with any significant change in the volatility of reduced-form innovations to real GDP growth. This is especially noteworthy for Japan and Finland, which, during those years, experienced a dramatic deceleration in potential output growth, and a prolonged and sizeable fall in the level of potential output, respectively. Fourth, in a few cases, severe financial crises appear to have been characterized by a temporary increase in the persistence of output growth. This is especially apparent for the U.S. during the recent crisis, and for Sweden in the early 99s. This implies that these crises have been associated with deep and prolonged recessions not only because (trivially) they have been characterized by a sequence of comparatively large negative s, but also because conditional on those s output growth has exhibited a systematic tendency to revert to the mean more slowly than under normal circumstances. On the other hand, once again, a temporary increase in output growth persistence has clearly not been a robust common feature of financial crises. Fifth, in the majority of cases impulse-response functions to the identified structural s do not exhibit any systematically different pattern compared to normal periods, thus suggesting that, historically, financial crises have not been associated with changes in the way structural disturbances propagate through the economy. Overall, international evidence for the post-wwii period therefore suggests that, beyond having been characterized by sequences of comparatively large contractionary s, severe financial crises have not exhibited a widespread pattern of systematic differences compared to standard macroeconomic fluctuations: as we pointed out, once controlling for the size of the s the only features that still stands out is a non-negligible, and sometimes significant negative impact on potential output dynamics.. Related literature In recent years, the persistent, negative impact on output of traumatic macroeconomic events such as banking and currency crises has been extensively documented in the literature. Based on data for 9 countries for the period 9-, Cerra and Saxena (8) showed how banking crises have uniformly been associated with large

5 and prolonged falls in output for all the country groups and geographical areas they considered. In particular, based on their full sample they estimated the change in output following a banking crisis to reach a plateau equaltoalmost-8percentatthe -year horizon. An intrinsic limitation of Cerra and Saxena s work, however, is that, being based on reduced-form methods, it can only estimate the average magnitude, within a specific group of countries, of the fall in output associated with a financial crisis, but it cannot identify either the nature of the s at the origin of the crises themselves, or the specific dimensions along which recessions associated with financial crises differ from normal recessions (apart from the size and duration of the fall in output). In their influential history of financial crises over the last several centuries, RR (9) documented systematic, large and persistent falls in real per capita GDP in the immediate aftermath of severe banking crises, equal, on average, to -9. per cent. As previouslypointedout,however,rr sentireworkisbasedontheanalysisoftheraw data, and, exactly as Cerra and Saxena (8), it cannot therefore provide answers to the three questions we posed at the beginning. Two papers have questioned, for the United States, RR s contention that financial crises have historically been fundamentally different from normal recessions. Based on both the analysis of the raw data, and simple reduced-form regressions, Lopez-Salido and Nelson () argued that one of RR s two previously mentioned key findings recessions associated with financial crises have been especially prolonged and drawnout does not hold for the post-wwii U.S.. A limitation of Lopez-Salido and Nelson s (henceforth, LSN) analysis, beyond being based on very simple regression methods, is however that they do not consider the Great Recession (which, at the time of their writing, was still ongoing). Indeed, as I discuss more extensively in Section below, for the post-wwii U.S. before the Great Recession I estimate, in line with LSN, essentially no change over time in the persistence of output growth, but crucially, and in line with RR, I identify a sizeable increase in persistence associated with the Great Recession. Bordo and Haubrich () s work is, both conceptually and methodologically, very close to LSN s, with the key difference being that it analyzes U.S. data since the second half of the XIX century. Their main conclusion is that [... ] recessions associated with financial crises are generally followed by rapid recoveries.. They identify three exceptions to this pattern: [...] the recovery from the Great Contraction in the 9s; the recovery after the recession of the early 99s and the present recovery [that is, the recovery from the Great Recession]. As for LSN, an intrinsic limitation of Bordo and Haubrich s analysis is that, being based on very simple regression methods, it cannot either identify structural s, or explore the specific dimensions along which recessions associated with financial crises have been idiosyncratic. Finally, Stock and Watson () focus on the 7-9 U.S. recession based on a fixed-coefficients dynamic factor model with variables. One of their key See RR (9, pages 9-, and in particular Figure.). 5

6 conclusions is that [...] although many of the events of the 7-9 collapse were unprecedented, their net effect was to produce macro s that were larger versions of s previously experienced, to which the economy responded in an historically predictable way. As previously pointed out, this is in line with most (but not all) of the results of the present work. An intrinsic limitation of either LSN (), Bordo and Haubrich (), or Stock and Watson () is their exclusive focus on the United States. It is indeed an open question to which extent the U.S. experience can be regarded as representative of the broader experience with financial crises worldwide. Along one conceptually related dimension the size of the unit root in GNP Cogley (99) showed for example that the U.S. should be regarded, when seen from an international perspective, as an outlier, with the size of the permanent component of output estimated to have been markedly smaller than for all other countries he considered. This implies that exploring this issue also based on data for countries other than the U.S. should be regarded, at the very least, as a robustness check on the results one obtains based on U.S. data. This is the fundamental reason why, in the spirit of RR (9), in the present work I have chosen to use data for as many countries as possible. The only other paper I am aware of to adopt an international perspective is Papell and Prodan (). Papell and Prodan (henceforth, PP) apply break tests within a univariate context in order to investigate whether severe recessions associated with financialcriseshaveapermanentimpactonoutput, and, if they don t, whether under these circumstances output returns to potential more slowly than during normal recessions. Overall, evidence for episodes comparable with the U.S. Great Recession suggests that, although output ultimately returns to the pre-crisis trend, on average slumps last about nine years. The paper is organised as follows. The next section describes the Bayesian methodology we use to estimate the time-varying parameters VARs with stochastic volatility, the identification strategy, and the methodology for computing the structural VAR s impact matrix. Section discusses the set of features which have historically been common across most severe financial crises, whereas Section focuses on individual financial crises idiosyncratic aspects. Section 5 concludes, and discusses directions for future research.

7 Methodology. A Bayesian time-varying parameter VAR with stochastic volatility In what follows we will work with the following time-varying parameters VAR(p) model: = () where the notation is obvious, and is defined as [ ],where is the logarithm of real GDP; is inflation, computed as the log-difference of the GDP deflator; is a short-term interest rate (in the case of the United States, the Federal Funds rate), which is quoted at a non-annualized rate in order to make its scale exactly comparable to that of inflation; 5 and is a stationary variable capturing the state of the business-cycle. For the United States, following Blanchard and Quah (989), is the civilian unemployment rate. For all other countries, for which the assumption of stationarity of the unemployment rate is questionable, is the consumption/output ratio. 7 The key rationale for including in the VAR either the unemployment rate or the consumption/output ratio is in order be able to effectively disentangle permanent andtransitoryoutputs. TheVARweestimatefortheU.S.isanexpanded version of the one originally estimated by Blanchard and Quah (989), which only included the log-difference of output and the unemployment rate, whereas the one we estimate for all other countries is an expanded version of one of the models estimated by Cochrane (99), which only included the log-difference of output and the log of the consumption/gdp ratio. 8 In both papers, the unemployment rate and the log of the consumption/output ratio, respectively, were included because of their strong informational content on the state of the business cycle and, therefore, on the 5 So,tobeclear,if is the relevant short-term rate with its scale such that, e.g., a ten per cent rate is represented as. is computed as =(+ /) -. Although the unemployment rate is bounded between and, so that, strictly speaking, it cannot be non-stationary, first, for all the countries considered herein (except the United States), bootstrapped p-values for Augmented Dickey-Fuller tests without trend clearly point towards the unemployment rate being I(). (These results are not reported here for reasons of space, but they are available from the author upoin request.) Second, even based on simple eyeball econometrics, for all countries except the U.S. the unemployment rate exhibits an obvious extent of permanent variation. In the Euro area, for example, it has increased from about - per cent in the early 97s to a new equilibrium around 9- per cent which has prevailed over the last decades. 7 Results based on the logarithm of the consumption/output ratio are near-identical to those based on the consumption/output ratio. These results are not reported here for reasons of space, but they are available from the author upon request. 8 As discussed by Cochrane (99, see in particular the Appendix) the benchmark model he estimates in the paper a cointegrated VAR for the log-differences of real GNP and real consumption can be re-cast in several alternative ways. One of them is as a VAR for the log-difference of real GNP and the logarithm of the consumption/output ratio. See also Cogley (5), who, conceptually in line with the present work, estimated a Bayesian time-varying parameters VAR for output growth and the logarithm of the consumption/output ratio. 7

8 transitory component of output. 9 For a complete description of the data and of their sources, see Appendix A. The overall sample periods are 95Q-Q for the United States, 9Q- Q for the Euro area, 97Q-Q for Finland, 955Q-Q for Japan, and 98Q-Q for Sweden. For all countries, however, we use the first years of data in order to compute the Bayesian priors, so that the effective sample periods start in 9Q for the United States, in 98Q for the Euro area and Finland, in 95Q for Japan, and in 99Q for Sweden. As it is customary in the literature on Bayesian time-varying parameters VARs, we set the lag order to p=. The VAR s time-varying parameters, collected in the vector, are postulated to evolve according to (, ) = ( ) (, ) () with ( ) being an indicator function rejecting unstable draws thus enforcing a stationarity constraint on the VAR and with (, ) given by = + () with ( ). The VAR s reduced-form innovations in () are postulated to be zero-mean normally distributed, with time-varying covariance matrix Ω which, following established practice, we factor as The time-varying matrices and are defined as: with the evolving as geometric random walks, Var( ) Ω = ( ) () (5) ln =ln + () For future reference, we define [, ]. Following Primiceri (5), we postulate the non-zero and non-one elements of the matrix which we collect in the vector [,,..., ] to evolve as driftless random walks, = +, (7) 9 As extensively discussed by Cochrane (99), the strong informational content of the consumption/output ratio for the transitory component of output is a direct consequence of the permanent income hypothesis under the assumption of rational expectations, and of frictionsless access to borrowing on the part of consumers. See e.g. Cogley and Sargent (), Cogley and Sargent (5), Primiceri (5), Benati (8), and Benati and Goodhart (). 8

9 andweassumethevector[,,, ] to be distributed as ( ),with = and = where is such that. As discussed by Primiceri (5), there are two justifications for assuming a block-diagonal structure for. First, parsimony, as the model is already quite heavily parameterized. Second, allowing for a completely generic correlation structure among different sources of uncertainty would preclude any structural interpretation of the innovations. Finally, following, again, Primiceri (5) we adopt the additional simplifying assumption of postulating a block-diagonal structure for, too namely Var ( )=Var ( )= (9) with Var( ), Var([ ] ),and Var([ ] ),thus implying that the non-zero and non-one elements of belonging to different rows evolve independently. As discussed in Primiceri (5, Appendix A.), this assumption drastically simplifies inference, as it allows to do Gibbs sampling on the non-zero and non-one elements of equation by equation.. Estimation We estimate ()-(9) via standard Bayesian methods. Appendix B discusses our choices for the priors which are standard and the Markov-Chain Monte Carlo (henceforth, MCMC) algorithm we use to simulate the posterior distribution of the hyperparameters and the states conditional on the data... Imposing constraints on the VAR s time-varying means The fact that both the unemployment rate and the consumption/output ratio are, by construction, bounded between and, whereas the short rate cannot take negative values, automatically implies that the corresponding time-varying means (that is: local equilibrium levels) implied by the VAR ought to satisfy those very same constraints at each point in time. In estimation we impose such constraints by rejecting each MCMC draw for which either (i) thevar smeanfortheunemploymentrate Primiceri (5, pp. -7). (8) 9

10 (or the consumption/output ratio) is outside the [, ] interval for at least one quarter, or (ii) the mean for the short rate is negative for at least one quarter. This way of imposing constraints on the VAR s means was used by Cogley and Sargent () to impose a non-negativity constraint on the mean for the short rate, and, from a conceptual point of view, it is akin to the previously discussed imposition of a stationarity constraint on the VAR on a period-by-period basis. An alternative way of imposing the relevant constraint on the means for the unemployment rate and the consumption/output ratio would have been to work with their logit transformations. For the consumption/output ratio this would have produced results near-identical to those reported below, since, for all countries we consider, the consumption/output ratio and its logit transformation once demeaned, and divided by their respective standard deviations are near-numerically identical to each other. Intuitively, this is due to the fact that, for values which are sufficiently far away from either or, the logit transformation is very close to being linear, and only for values close to the two boundaries it becomes highly non-linear. For the U.S. unemployment rate, however, results would likely have been different, because, even when standardized, the unemployment rate and its logit transformation differ from each other to a non-negligible extent for several quarters. (This is due to the previously mentioned high non-linearity of the logit transformation for values sufficiently close to either or.) In particular and crucially for the present purposes during the Great Recession the logit of the U.S. unemployment rate increased proportionally less than the unemployment rate itself, which would have misleadingly pointed towards a milder recession than it actually was the case. As a consequence, had we worked, for the United States, with the logit of the unemployment rate, we would likely have under-estimated the extent of the transitory fluctuation of output during that episode, and we would therefore have likely over-estimated the relevance of permanent output s.. Identification.. The permanent output We identify four s. Following Blanchard and Quah (989), the first is defined as the only one exerting a permanent impact on log output, and it is identified via a long-run restriction. Giventhekeyroleplayedbysuchpermanent-transitorydecompositioninour analysis, it is important to have some corroborating evidence on the reliability of the So, to be clear, given the sequential nature of the MCMC algorithm used herein, if all of the relevant constraints are satisfied for the j -th draw we move to(j +)-th draw. Otherwise, we reject the j -th draw and we keep drawing until we get a draw which satisfies all of the relevant constraints on the VAR s means for each quarter. Only at that point we move to (j +)-th draw. Indeed, Cogley and Sargent () impose such a constraint on the unemployment rate s mean by working with the logit of the unemployment rate.

11 Blanchard-Quah identification scheme when applied to our time-varying parameters VARs. To put it differently, what are the properties of the identified permanent and transitory output s? Do they look reasonable when seen from the perspective of reliable information extraneous to the VARs? The left-hand side panel of Figure provides informal evidence on this, by showing, for the United States, the median and the one-standard deviation percentiles of the posterior distribution of the estimated transitory components of log real GDP, together with the output gap estimate implied by the Congressional Budget Office (henceforth, CBO) estimate of potential GDP. The CBO and VAR-implied estimates of the output gap strongly co-move and are, most of the time, numerically quite close. The main exceptions are the second half of the 9s, when the CBO estimate, peaking at around per cent, is significantly larger than the one generated by the VAR, and the period following the collapse of Lehman Brothers, with the VAR pointing towards a smaller transitory fluctuation (and therefore, a correspondingly larger permanent one) than the one implied by the CBO estimate. As we will discuss in Section, indeed, a non-negligible impact on potential output dynamics appears to have been, historically, a feature common to all the severe financial crises analyzed herein. Overall, we regard the evidence reported in the left-hand side panel of Figure as reassuring, as, for the only country for which there exists an authoritative output gap estimate which can be used as a benchmark, 5 the transitory component of output generated by the Blanchard-Quah decomposition applied to the time-varying VAR appears as entirely reasonable. The other two panels of Figure report estimates of the transitory component of output for Japan and the Euro area, whereas the second column of Figure shows the corresponding estimates for Finland and Sweden. Overall, VAR-generated estimates of the output gap appear as entirely reasonable, and they accord well with post- WWII macroeconomic history. For Japan, for example, we estimate (based on median estimates) a sizeable output gap, equal to about per cent, at the peak of the bubble, at the end of the 98s-beginning of the 99s, and, in line with the conventional wisdom of a prolonged recession during subsequent years, a consistently negative output gap for most of the quarters between the early 99s and the beginning of the financial crisis. For the Euro area, the output gap is estimated to have decreased, during the financial crisis, from a peak of. per cent in 8Q to a trough of -. per cent in Q, and to have subsequently increased, reaching a value of -. per So, to be clear, what is labelled in the Figure as the CBO estimate of the output gap has been computed as the percentage difference between GDPC9 ( Real Gross Domestic Product, Decimal, Seasonally Adjusted Annual Rate, Billions of Chained 5 Dollars ) and GDPPOT ( Real Potential Gross Domestic Product, U.S. Congress: Congressional Budget Office, Billions of Chained 5 Dollars ). It is important to stress that both GDPC9 and GDPPOT are expressed in the same units, and so they are exactly comparable. 5 For the Euro area, the only output gap estimate we are aware of is the European Central Bank s own internal estimate, which, however, is not disseminated externally. For other countries, we are not aware of any available official estimate.

12 cent at the end of the sample, in Q... The three transitory output s The other three s which, by construction, only exert a transitory impact on output are identified based on the standard set of sign restrictions reported in the following table. Shock: Variable: Output growth Inflation Unemployment rate, or consumption/output ratio Short rate?? = left unconstrained An interest rate ( )isidentified based on the restriction that it exerts a non-negative impact on both the interest rate and either the unemployment rate or the consumption/output ratio, and a non-positive impact on inflation and output growth. A demand non-interest rate ( ) is postulated to have a non-negative impact on either output growth, inflation, and the interest rate, and a non-positive impact on either the unemployment rate or the consumption/output ratio. Finally, a transitory supply ( ) is disentangled from the other two because it is postulated to be the only inducing a negative co-movement between inflation and output growth, and a positive co-movement between inflation and either the unemployment rate or the consumption/output ratio. In what follows we impose these sign restrictions only on impact. The reason for doing this is that, as stressed by Canova and Paustian (), whereas sign restrictions on impact are, in general, robust in the specific sense that they hold for the vast majority of sub-classes within a specific class of DSGE models, and for the vast majority of plausible parameters configurations restrictions at longer horizons are instead, as they put it, whimsical, meaning that they are hard to pin down, and in general, they are not robust across sub-classes of models, and for alternative plausible parameters configurations. One obvious limitation of imposing the sign restrictions only on impact is that we are here using a comparatively limited amount of information in order to achieve identification. As a consequence, our results necessarily end up being less sharp than they could have been had we been reasonably confident about imposing a specific pattern of sign restrictions at horizons greater than zero. This compounds a well-known limitation of sign restrictions which has been extensively discussed by Fry and Pagan (7): as these authors stress, sign restrictions are intrinsically weak information, since they are based on the notion of uniquely imposing a specific pattern of signs on the IRFs. The rationale behind our decision of imposing sign restrictions only on impact is that it is better to

13 .. Rationale behind the postulated sign restrictions Transitory output s should logically be expected to move output and the unemployment rate in opposite directions for evidence supporting this position, see the IRFs of output and the unemplyment rate to a transitory output estimated by Blanchard and Quah (989). As for the response of the consumption/output ratio, the negative co-movement of this variable with output in response to transitory output s is a direct consequence of the permanent income hypothesis (on this, see the extensive discussion in Cochrane, 99). 7 This implies that, in discussing the rationales behind the postulated sign restrictions, we can ignore, from now on, the unemployment rate and the consumption/output ratio, and we can exclusively focus on output growth, inflation, and the short rate. The sign restrictions reported in the previous table are the same used, e.g., by Benati (8), Benati and Goodhart (), and Benati and Lubik (), and can be motivated in two diffent ways. First, they naturally arise from a simple aggregate demand-aggregate supply framework. Second, they are the same as the robust sign restrictions reported by Canova and Paustian () in their Table for their benchmark DSGE model featuring sticky prices, sticky wages, and several standard frictions (see the column labelled as M ). Specifically, the signs we impose for the impacts of on the short rate, inflation, and output growth are the same as those for Canova and Paustian s monetary, whereas the signs we impose for are the same as those for their taste. Finally, as for, the opposite pattern of signs it induces on impact on inflation and output growth holds for either of Canova and Paustian s markup and technology s. 8.. What about credit market disturbances? As I also discuss in Section.5 below, credit market disturbances should logically be expected to end up being classified, within the present taxonomy of s, as demand non-interest rate s. 9 A positive to credit spreads, for example, impose a limited amount of information about which we can be reasonably confident than a greater amount of information about which we have limited confidence. 7 Whereas a permanent output moves output and consumption in lockstep, thus leaving their ratio unaffected, a transitory only impacts upon output, and it instead leaves consumption unaffected. As a result, transitory output s, whatever their specific nature and origin, necessarily induce a negative co-movement between output and the consumption/output ratio. 8 Canova and Paustian s Table also reports robust sign restrictions for the impacts of the markup and technology s on the short rate. However, since such restriction is not necessary in order to disentangle from the other two transitory output s (as we pointed out, is the only inducing a negative co-movement between inflation and outpout growth), we have preferred not to impose it. 9 On the other hand, credit market disturbances associated with financial institutions going out of business, or even with a permanent shrinking of the size of the financial sector (which appears to have been the case, e.g., in the United Kingdom during the recent financial crisis) would have a permanent impact on output, and would therefore end up being classified as permanent output

14 should be expected to have a non-positive impact on both inflation and output growth. As for the short rate, the impact should also be expected to be non-positive, based on the rationale that, facing such a, the central bank may choose to to do nothing in which case the short rate would remain unchanged or (which is more likely, especially in the light of the recent financial crisis) it may decide to counter the contractionary impact of the increase in credit spreads by cutting the policy rate. On the other hand, it is highly implausible that, in the face of such a, the monetary authority may somehow decide to compound its contractionary impact by increasing the policy rate. This implies that, e.g., if credit market disturbances played a dominant role during financial crises in terms of the fractions of the forecast error variance of macroeconomic variables explained by such s this should be expected to show up in an increase in the importance of demand non-policy s. On the other hand, if, once controlling for changes in the volatility of all s, features pertaining to demand non-policy s did not exhibit, during financial crises, significant differences compared to standard busines cycle fluctuations, this would suggest that, under this specific dimension, financial crises differ from normal periods only in terms of the greater volatility of s.. Computing the VAR s structural impact matrix For each quarter, and for each draw from the ergodic distribution, we compute the time-varying structural impact matrix, by combining the methodology proposed by Rubio-Ramirez, Waggoner, and Zha (5) for imposing sign restrictions and the procedure proposed by Gali and Gambetti (9) to impose long-run restrictions within a time-varying parameters VAR context. Specifically, let Ω = be the eigenvalue-eigenvector decomposition of the VAR s time-varying covariance matrix Ω,andlet. We draw an matrix,, fromthe (, ) distribution, we take the decomposition of that is, we compute matrices and such that = and we compute the time-varying structural impact matrix as =. Following Gali and Gambetti (9, Section II), we then compute a local approximation to the matrix of the cumulative impulse-response functions (henceforth, IRFs) to the VAR s structural s as =[ ] {z } () s. See at The only difference between Gali and Gambetti (9, Section II) and the present work is that they compute the local approximation to the matrix of the cumulative IRFs based on the companion form of the VAR, whereas we compute it directly based on the VAR itself.

15 where is the identity matrix. We then rotate the matrix of the cumulative impulse-response functions via an appropriate Householder matrix in order to introduce zeros in all of the first row of that is, the row corresponding to log real GDP except for the (,) entry, so that the firstrowoftheresultinglocal approximation to the matrix of the cumulative impulse-response functions, = = = () is given by =[ ( ) ],with ( ) beingavectorof( -) zeros, and being a non-zero entry. This imples that the first is the only one exerting a long-run impact on the level of log output. If the resulting structural impact matrix = satisfies all the sign restrictions we keep it, otherwise we discard it and we repeat the procedure until we obtainanimpactmatrixwhichsatisfies both the sign restrictions and the long-run restriction at the same time..5 A possible limitation of the present work As extensively discussed by DelNegro (), time-varying parameters VARs suffer from a curse of dimensionality, in the specific sense that imposing upon them a stationarity constraint on a period-by-period basis becomes progressively more difficult as the number of variables entering the VAR increases. This is the key reason why, in the present work, we have chosen to limit ourselves to output growth, inflation, a short rate, and either the unemployment rate or the consumption/output ratio. Although the rationale for doing so is a compelling one, our analysis suffers from the potential drawback that, by eschewing variables such as credit spreads, we are disregarding information which may be important for the purpose of correctly identifying the dimensions along which recessions associated with financial crises differ from normal business-cycle fluctuations. Although potentially relevant in principle, this problem should be put into perspective. As we previously discussed in Section.., within the present taxonomy of s credit market disturbances should be expected to end up being classified as demand non-interest rate s, thus implying that any anomalous feature of recessions associated with financial crises pertaining to such disturbances should manifest itself in a corresponding anomalous feature for demand s. This means that the present analysis should be perfectly capable of providing answers to the questions listed in the three bullet points in the Introduction. We compute the Householder matrix via Algorithm 5.5. of Golub and VanLoan (99). Within the present context, imposing such a constraint is necessary in order to be able to (i) meaningfully compute IRFs to the structural s, and (ii) Fourier-transform the VAR at each point in time. Further, as argued by Cogley and Sargent (), the notion that an advanced country s economy might be non-stationary should be rejected on logical grounds. Based on my own experience, in practice it is not possible to estimate time-varying VARs with more than five variables, and indeed we are not aware of anybody who has ever been able to do that. 5

16 Ideally, we would have preferred to also include variables containing direct information on credit market developments in terms of both prices and quantities (e.g., credit spreads and credit supply growth). This, however, would have required us toworkwithatleasttwoadditionalseries,withtheresultthatwewouldnothave been able to estimate the VAR. 5 Although unwillingly, we have therefore resorted to working with the present specification, which in our own view effectively balances the ineludible constraint of computational feasibility with the ability to provide an answer to the questions we are exploring.. Identifying severe financial crises The financial crises we classify as severe are those which have been thus identified by RR (9), that is, for the countries and sample period considered herein, Finland and Sweden in the early 99s; Japan s bursting of the credit and real estate bubble in those very same years; and, for the United States and the Euro area, the recent financial crisis. Altough neither Japan nor Sweden have been at the epicenter of the recent crisis so that they have mostly been affected by it via trade channels, but they have not experienced severe stresses in their banking systems we have chosen to also report results for these two countries by way of comparison. The fact that, during the recent crisis, we identify a negative impact on potential output dynamics also for Japan and Sweden provides prima facie evidence that such an impact which, as we discuss in the next section, is the only feature we identify to have been common across all financial crises analyzed herein does not pertain to financial crises per se, and, most likely, is simply the consequence of the extraordinary depth of the recessions associated to such crises. A Feature Common to All Severe Financial Crises: The Negative Impact On Potential Output Dynamics Figure shows log real GDP, together with the median and the one-standard deviation percentiles of the posterior distribution of its estimated permanent component, for the U.S., the Euro area, and Japan, whereas the left-hand side column of Figure shows the same objects for Finland and Sweden. Even the simple eyeball metric clearly suggests that all severe financial crises considered herein have been characterized by a non-negligible negative impact on potential output dynamics, which in two 5 Since one of our key results pertains to the impact of severe financial crises on potential output dynamics, being able to effectively identify a permanent output is here of paramount importance. This implies that, beyond output growth, inflation, and the short rate which provide a minimal statistical summary of any economy, and should therefore be there we need a fourth variable with a strong informational content on the state of the business cycle.

17 cases Japan, following the collapse of the asset prices bubble of the second half of the 98s, and Finland, in the early 99s appears to have been nothing short of dramatic. Figure provides statistical evidence on this, by showing the posterior distributions of the difference between the annualized change in the estimated permanent component of GDP from the temporary peak to the temporary trough associated with the crisis, and the annualized change in the same object during the previous five years. 7 The only exception is Japan s crisis of the early 99s, for which we show the posterior distribution of the difference between the annualized change in potential output during the period following that crisis and up to the beginning of the most recent crisis, and the annualized change in potential output during the entire period before the first crisis. We do this in order to capture and characterize the obvious decrease in the rate of growth of Japan s potential output following the collapse of the bubble of the second half of the 98s. Finally, as for Sweden s crisis of the first half of the 99s, the sample period is too short to allow us to make any comparison between potential output dynamics before and during the crisis, and so we simply ignore it. The visual evidence from the bottom-left panel of Figure, however, suggests that, around the time of that crisis, potential output growth had been temporarily depressed compared to the pace of growth potential output had exhibited during the subsequent period, and up to the most recent crisis. Overall, the evidence reported in Figure paints a consistent picture, with all the financial crises analyzed herein having systematically been associated with either a temporary or in the case of Japan s crisis of the early 99s a permanent deterioration in the evolution of potential output. Two things, in particular, ought to be stressed about the evidence reported in the Figure. First, statistical significance is very strong, with the fraction of draws from the posterior distribution for which the impact of the financial crisis on potential output dynamics is estimated to have been negative being greater than, or equal to 99 per cent for all countries and all crises, with the single exception of the United States during the most recent episode, for which only 9.9 per cent of the draws are associated with a temporary deterioration of potential output growth. Second, the estimated magnitudes are definitely not negligible, with median estimates of the impacts, in terms of the (temporary) change in the annualized growth rate of potential output, ranging from -. per cent for Sweden during the recent Temporary peaks and troughs in potential output have been identifiedbasedonthemedian estimates of the permanent component of log real GDP. The dates of the peaks and troughs are the following: for the United States, 8Q and 9Q; for the Euro area, 8Q and 9Q; for Japan, 8Q and 9Q for the most recent recession, whereas for the previous episode we compare the period from the beginning of the sample up to 99Q and the period from 99Q to 7Q; for Finland, 99Q and 99Q for the first crisis, and 8Q and 9Q for the most recent; and for Sweden, 8Q and 9Q. 7 We take, as a benchmark, the previous five years, rather than a longer period, in order to take into account of the possibility of slow, secular changes in the rate of growth of potential output. 7

18 crisis, to a remarkable -. per cent for Finland during the three years between the potential output peak of 99Q and its subsequent trough of 99Q, when (based on median estimates) potential GDP decreased by 7. per cent. The results shown in Figure complement, and expand upon, RR s findings on the macroeconomic impact of financial crises. Not only, as they have documented, such episodes have consistently been characterized by dramatic falls in actual output: our findings point towards a similarly consistent and sizeable impact on potential output dynamics. It is also to be stressed that these results are by no means incompatible with PP s () finding, based on break tests applied within a univariate context, that in most cases, following a financial crisis, output ultimately returns to the precrisis trend. As PP point out, the average slump associated with such crises lasts about nine years, so the fact that that the pre-crisis potential output trend reasserts itself over such a long horizon is not incompatible with the notion that, at much shorter horizons, potential output dynamics is negatively affected to a non-negligible extent. In the light of Athanasios Orphanides work on output gap mismeasurement, and its role in inducing policy mistakes which contributed to igniting the U.S. Great Inflation, 8 our findings have an obvious policy relevance. To the extent that policymakers interpreted the fall in actual output associated with financial crises as entirely due to a decrease in the transitory component of output, they would over-estimate the size of the output gap, and would therefore end up over-estimating the disinflationary pressures in the economy.. Are financial crises special, or is it just the depth of the recession? Although a negative impact on potential output dynamics appears to have been a feature common to all crises and all countries analyzed herein, this does not necessarily imply that such a feature is specific to financial crises per se:an alternative, plausible interpretation in indeed that such an impact is simply the consequence of the extraordinary depth of the recessions associated with severe financial crises, so that such contractions end up blasting away pieces of the economy, thus negatively impacting potential output dynamics. Prima facie evidence in favor of such a position is provided by the fact that, during the recent crisis, both Japan and Sweden neither of which has been at the epicenter of the crisis, so that they have mostly been affected via trade channels, but they have not experienced severe stresses in their banking systems have nonetheless exhibited non-negligible, and highly statistically significant negative impacts on potential output dynamics. The case of Sweden is, under this respect, especially interesting, since (i) followingthefinancial crisis of the early 99s, its financial system 8 See in particular Orphanides (), Orphanides (b), Orphanides (), and Orphanides (a). 8

19 had been largely revamped and made significantly more resilient, and, most likely as aresultofthis,(ii) as we write (July 5th, ) it has not exhibited, so far, any problem comparable to those which have instead bedevilled (first and foremost) the United States and the Euro area. By the same token, Finland which belongs to the Euro area, but, apart from having experienced a deep recession, it has been, so far, only marginally affected by the financial crisis has exhibited a dramatic and highly statistically significant negative impact on potential output dynamics. The experience of these three countries during the recent crisis therefore naturally suggests that the negative impact on potential output dynamics we previously identified has nothing to do with financial crises per se, and it rather likely originates from the fact that as extensively documented by RR (9), such crises have been characterized by extremely deep recessions. Evidence for the United States is compatible with this position. Before the recent crisis, the deepest post-wwii contraction had been the Volcker recession, which had been characterized by a trough in annual real GDP growth equal to -.75 per cent (in 98Q), compared to the recent crisis trough of -5.5 per cent in 9Q. And in fact, compatible with our conjecture, the Volcker recession had indeed been characterized by a comparatively mild negative impact on potential output dynamics. By the same token, the evidence reported in the second panel of Figure shows how, in the Euro area, the comparatively mild recession of the first half of the 99s was associated with a temporary deceleration in the rate of growth of potential output, but no decrease in its level. The main exception to the argument that financial crises are not special, in terms of their negative impact on potential output dynamics, is provided by Japan s crisis of the early 99s, which was associated with a sizeable decrease in trend output growth, and is therefore, in a real sense, in a league of its own. Let s now turn to a discussion of the features which have been idiosyncratic to individual crises. Features Idiosyncratic to Individual Financial Crises. A temporary increase in macroeconomic volatility Figure 5 reports evidence on changes over time in macroeconomic volatility, by showing the median and the one-standard deviation percentiles of the posterior distributions of the standard deviation of reduced-form innovations to the log-difference of real GDP (in the top row), and of the log-determinant of Ω (in the bottom row), which, following Cogley and Sargent (), we interpret as a measure of the VAR s total prediction variance, that is, of the total amount of noise hitting the system at each point in time. For both the United States and the Euro area, the recent crisis has been charac- 9

20 terized by sizeable increases in both features. 9 This is especially apparent for the Euro area, where the standard deviation of reduced-form s to real GDP growth has temporarily increased reaching (marginally) the highest value in the entire sample period, whereas the VAR s total prediction variance has interrupted the long-run secular decline it had been experiencing since the mid-97s, returning, in the aftermath of the collapse of Lehman Brothers, to values it took in the early 99s. The experience of the other three countries, however, shows that sizeable, temporary increases in macroeconomic volatility are not a robust common feature across all financial crisis. First, for neither Japan, Finland, nor Sweden the financial crises of the early 99s were associated with any significant change in the standard deviations of reduced-form s to GDP growth. This is especially noteworthy for Japan and Finland, which, during those years, experienced a dramatic deceleration in potential output growth, and a prolonged and sizeable fall in the level of potential output, respectively. Second, the evolution of the VAR s total prediction variance is even more intriguing: whereas Finland s crisis of the early 99s was characterized by a temporary increase in the log-determinant of Ω, the corresponding crises in Japan and Sweden were associated with no significant change, and a sizeable decrease, respectively.. The persistence of output growth By the same token, the evidence reported in Figure shows that a temporary increase in the persistence of real GDP growth which, intuitively, we might be induced to expect, in the light of RR s finding about the length and depth of downturns associated with financial crises has not been, in fact, a robust common feature of financial crises. The top row shows the median and the one-standard deviation percentiles of the posterior distribution of the fraction of variance of real GDP growth pertaining to the frequency band [, ], which, at the quarterly frequency, is associated with fluctuations with a frequency of oscillation slower than eight years. The bottom row shows, for each quarter t, the fraction of draws from the posterior distribution for which the fraction of variance of real GDP growth pertaining to the frequency band [, ] in a quarter of reference is greater than it is in quarter t. 9 For either the U.S. or the Euro area, both the standard deviation of reduced-form innovations to the log-difference of real GDP, and the log-determinant of Ω, clearly appear to have started increasing well before the beginning of the crisis, in August 7. The simplest and most logical explanation for this result is that the estimates produced by the Gibbs sampler are, by construction, two-sided, and in the case of processess experiencing sharp breaks they therefore tend to inevitably smooth the break, thus mixing the future with the past, and giving the misleading impression that the change took place before it actually did. The quarters of reference are those corresponding to the maximum values taken by the medians showninthefirst row. Evidence based on the normalized spectrum of output growth at = is very similar to that shown in Figure. These results are not shown here for reasons of space, but they are available

21 Only for two countries severe financial crises clearly stand out, compared to standard business-cycle fluctuations, for being characterized by sizeable, temporary increases in the persistence of output growth. The first is the United States, where, during the recent crisis, the fraction of variance of real GDP growth pertaining to the frequency band [, ] increased (based on median estimates) from 7 per cent in 7Q to a peak of per cent in the quarter following the collapse of Lehman Brothers, 8Q. Since then it has been steadily decreasing, reaching about per cent in the last quarter of the sample. The fraction of draws for which the fraction of variance of real GDP growth pertaining to the band [, ] in 8Q is greater than in quarter t paints a similar picture, falling from 8 per cent in 7Q to slightly below 5% in 9Q, and then rebounding over subsequent quarters. The second country is Sweden, where, during the crisis of the early 99s, the fraction of variance of real GDP growth associated with the low frequencies experienced a similar, but milder temporary increase, and the fraction of draws, swiftly decreased from 78 per cent in 99Q to 5% in the quarter of reference, 99Q, and then quickly rebounded, reaching 9 per cent in 99Q. This implies that these two episodes have been associated with deep and prolonged recessions not only because (trivially) they have been characterized by comparatively large contractionary s, but also because conditional on those s output growth has exhibited a systematic tendency to revert to the mean more slowly than under normal circumstances. For all other countries, on the other hand, evidence does not support the notion that a temporary increase in the persistence of real GDP growth is a robust feature of severe financial crises. For Japan, for example, the dramatic crisis of the early 99s was not associated with any perceptible change in output growth persistence, whereas the recent crisis has been characterized by a comparatively mild increase. By the same token, Finland s crisis of the early 99s was not accompanied by any significant change in output growth persistence. Evidence for the Euro area is more complex. On the one hand, following the collapse of Lehman Brothers output growth persistence appears indeed to have increased, although a proper interpretation of the evidence is complicated by the fact that years immediately preceding the crisis had been characterized by a decrease roughly of the same magnitude. The counterargument to the notion that a large increase in output growth persistence is a typical feature of financial crises is provided by the recession of the early 98s, which was characterized by an extent of real GDP growth persistence just marginally lower than the one associated to the recent crisis.. The role of individual structural disturbances Let s now turn to the role played by individual structural disturbances. The next two sub-sections explore the specific sign pattern exhibited by such disturbances within the context of individual financial crises, and the counterfactual paths the economy from the author upon request.

22 would have travelled had individual s been absent... The sign pattern of structural s Figures 7 to 9 show, for either country, and for either quarter, the fractions of draws from the posterior distribution for which the four identified structural s are estimated to have been positive. Within each panel, the red broken line represents theaveragevalue takenbysuch afractionover the entire sample period. The reason for reporting this object is in order to be able to more precisely characterize to which extent specific periods have been anomalous compared to the normal pattern of s signs for the entire sample period. Starting from the United States, the years immediately preceding the recent crisis had been characterized by predominantly expansionary demand non-policy s, and by predominantly contractionary supply s (either permanent or transitory), and, marginally, by predominantly expansionary monetary policy s. Results for the permanent supply s between the beginning of the century and the onset of the financial crisis resonate with the well-known productivity slowdown which took place during those years, 5 whereas the weak evidence of expansionary monetary policy s is incompatible, e.g., with John Taylor s criticism of the Federal Reserve s monetary policy during those years, and is instead compatible with the FED s often-stated position that, in the run-up to crisis, its monetary policy had not been especially expansionary. 7 Particularly interesting is the reversal of the sign pattern for the demand non-policy s, which are estimated to have been most likely contractionary between the early s and mid-, and most likely expansionary between mid- and the begining of the crisis. Since, as previously On the other hand, we eschew the issue of the volatility of individual structural s since, in order to identify such volatilities, we would need to impose a set of normalizations on the impacts of the s (see, e.g., Gambetti, Pappa, and Canova ())). Since any such normalization is essentially arbitrary, we prefer to simply ignore the issue, rather than presenting results we would regard as questionable. It is to be noticed how such fractions fluctuate quite significantly over the interval [, ], with the single exception of the permanent output. Intuitively, this is due to the fact that potential output is estimated to have evolved very smoothly, so that the fraction of draws for which permanent output s are estimated to have been positive has consistently remained close to one half. Although such averages have almost uniformly been near-identical to.5, in a few instances they have clearly been either slightly higher or slightly lower, thus suggesting that the corresponding s have been, on average, either expansionary or contractionary. This has marginally been the case for the potential output and monetary policy s for the Euro area, which are estimated to have been, on average, mostly contractionary; and of the demand non-policy s for the U.S., Japan, and Sweden, which have been, instead, mostly expansionary. 5 E.g., the annual rate of growth of output per hour in the non-farm business sector decreased from an average value of. per cent over the period Q-Q, to an average value of. per cent over the period 5Q-7Q. See e.g. Taylor (9). 7 [Here put references: Bernanke s speech at the AEA meetings, and the work of Kiley et al.]

23 discussed, credit market s should be expected to end up being classified, within the present taxonomy, as demand non-policy s, this suggests that such s played most likely an expansionary role in the run-up to the crisis. Since August 7, on the other hand, the sign pattern has once again reversed, and, consistent with what we would expect just based on (say) reading the financial press, it has been most likely negative. Potential output s have been most likely negative between the beginning of the crisis and 9Q, but over subsequent quarters the sign pattern has most likely switched to positive. 8 Finally, monetary policy s have been most likely expansionary between mid-7 and the collapse of Lehman Brothers, but they have been most likely contractionary during subsequent quarters. Interestingly, Gali, Smets, and Wouters () obtain the same result based on an estimated DSGE model, and interpret it as the logical consequence of the zero lower bound having been binding during that period. For other countries, the structural s sign patterns do not exhibit any consistent similarity compared to the United States. For the Euro area, for example, neither the monetary policy nor the demand non-policy s display any clear-cut consistent sign pattern during the recent crisis, whereas, in line with the discussion of Section, the negative impact on potential output dynamics is largely concentrated in the quarters between the beginning of the crisis and the collapse of Lehman Brothers. Consistent with the U.S. recent crisis, Japan s crisis of the early 99s had been characterized by most likely expansionary demand non-policy s in the run-up to the crisis, and by most likely contractionary s over the subsequent period. Monetary policy s had been most likely expansionary around the mid-99s and in the run-up to the recent crisis, but other than that they have not exhibited any clear-cut sign pattern. Likewise, transitory supply s have been most likely expansionary in the run-up to both crises, but again, other than that, they have not displayed any consistent pattern. Finally, given the dramatic slowdown in Japan s potential output growth starting in the early 99s, the lack of any consistent sign pattern in permanent output s over the entire sample period may appear, at first sight, as puzzling. It is important to stress, however, that given the permanent nature of the slowdown in real GDP growth, the time-varying VAR has captured it by means of a permanent decrease in the VAR-generated unconditional mean for the log-difference of GDP. To put it differently, since the slowdown in output growth has been permanent, the VAR has allocated it its time-varying unconditional mean vector, rather than to the permanent output s. Turning to Finland and Sweden s crises of the early 99s, the pattern for demand non-policy s is in line with that for Japan s crisis during the same period, and for the United States during the recent crisis: for Finland, s had been most likely positive in the run-up to the crisis, and negative therefore; for Sweden, on the other hand, we can t say anything about the run-up to the crisis, which was before the start of the sample period, but 8 Once again, this resonates with the evolution of the rate of growth of output per hour in the non-farm business sector, which, likewise, exhibited a dramatic peak in 9Q.

24 following the outbreak of the crisis s had most likely been negative. For both countries, monetary policy s are were most likely positive during the quarters immediately following the outbreak of that crisis, whereas potential output s had been most likely negative for Finland, but, quite strikingly, most likely positive for Sweden. Finally, during the recent crisis, the most clear-cut sign pattern pertains potential output s in Finland, and, to a slightly lesser extent, in Sweden, which, as expected, have been almost uniformly negative, whereas other s do not exhobit any clear pattern. Overall, the sign pattern of structural s in the run-up to, and during financial crises has no exhibited any clear-cut consistent behavior. It is to be noticed, however, that in most cases with the most notable exception being the Euro area during the recent crisis demand non-policy s, which should here be expected to capture credit market disturbances, did indeed display a consistent pattern, expansionary in the run-up to the crisis, and contractionary following its outbreak. So, although, strictly speaking, our results do not point towards a universal, consistent pattern for structural s signs, for demand non-policy s we come quite close to that... Counterfactuals shutting off individual structural s 9 An alternative way of assessing the role played by individual structural disturbances is to explore what the time path of the endogenous variables would have been had one of those disturbances been absent. Figure to 7 provide evidence on this, by showing, for each country, and for all the financial crises considered herein, the actual paths of annual real GDP growth, annual inflation, and the short rate, together with the median and the th and 8th percentiles of the posterior distributions of the counterfactual paths obtained by setting to zero one structural at a time. For reasons of space, in what follows I will mostly discuss results for real GDP growth, and I will only mention results for other variables when I deem them to be of particular interest. Starting from the United States, the results shown in the first column of Figure highlight how, within the context of the recent crisis, all s have played an overall contractionary role, as setting to zero either of them systematically generates a higher counterfactual path for real GDP growth. At the same time, it is apparent how 9 For reasons of space we do not discuss results for the fractions of forecast error variance of individual series at the various horizons attributable to individual structural disturbances. Unsurprisingly, these results are, conceptually, essentially the same as those discussed in this sub.section. All of these additional results are available from the author upon request. So, to be clear, each counterfactual has been computed by taking as given the evolution of the log-differences of real GDP and the GDP deflator, of the short rate, and of either the unemployment rate or the consumption-output ratio up to a quarter -. Then, starting from quarter, marking the beginning of the financial crisis of interest, we re-run history by setting to zero one structural at a time. As for the recent crisis, is set to 7Q for all countries. As for the financial crises of the early 99s, on the other hand, is set to 989Q for Finland and Japan, and to 99Q for Sweden.

25 setting to zero demand non-policy s only produces a slightly higher counterfactual path thus pointing towards a comparatively limited role played by such s within the context of the recen crisis whereas, by the same token, the role played by transitory supply s has been only marginally more important. On the other hand, evidence suggests that permanent output s and monetary policy s have played a dominant, and roughly equal role, with the counterfactual paths for real GDP growth being significantly higher than the actual, historical path. Consistent with the discussion of the previous sub-section, the fact that killing off monetary policy s generates a higher counterfactual path for real GDP growth derives from the fact that, within the context of the Great Recession, the zero lower bound has been binding. In turn, this implies that, for the most recent period, the structural VAR necessarily ends up estimating positive monetary policy s, so that, setting them to zero in the counterfactual, automatically produces a higher paths for output growth. Interestingly, as far as inflation is concerned the only counterfactual to produce materially different results specifically, a higher inflation path is the one in which the monetary policy is set to zero. Turning to the Euro area,the key finding emerging from Figure is the dominant role played by the permanent output within the context of the recent crisis. Whereas setting to zero either of the other three s produces counterfactual paths for the three series of interest which are only marginally different from the actual, historical paths, killing off the permanent output generates significantly higher counterfactual paths for all series. In particular, based on median estimates real GDP growth would have experienced a trough equal to -. per cent in 9Q, as opposed to an actual trough of -5.5 per cent in the previous quaretr, whereas inflation would have experienced a modest fall, with a counterfactual trough equal to.5 per cent, compared to the actual trough of. per cent. As a result of this, the short rate would have remained systematically higher, and instead of rapidly falling towards, and then remaining close to zero following the collapse of Lehman Brothers, it would have reached a trough of. per cent in Q, and it would have then edged slightly upwards over subsequent quarters. In spite of the dramatic deceleration in real GDP growth following Japan s crisis of the early 99s, killing off permanent output s would have produced an only marginally higher counterfactual path for output growth. This originates from the fact that, as we pointed out in the previous sub-section, the permanent decrease in output growth has been captured by the VAR by means of a permanent decrease in the VAR-implied time-varying mean for this series, thus automatically minimizing the role played by permanent output s. By the same token, setting to zero either of the other s would have produced the same result. As for the recent crisis, the most interesting result is that, in spite of the previously discussed temporary decrease in potential output associated with the crisis, killing off permanent output s does not produce a huge increase in real GDP growth. This originates from the fact that, as it is apparent from Figure, the peak-to-trough decrease in actual output during 5

26 the crisis was significantly larger than the corresponding peak-to-trough decrease in potential output, which implies that, different from the Euro area, transitory output s played a dominant role. As the remaining three panels of the first column of Figure show, in fact the key role was played by transitory supply s, whereas the two demand-side s role has been essentially negligible. Permanent output s played on the contrary a dominant, and dramatic role within the context of Finland s crisis of the early 99s. Consistent with the potential output estimates reported in Figure, killing off such s would have generated a significantly higher counterfactual path for output growth. In particular, based on median estimates, first, real GDP growth would have reached a trough of -.9 per cent in 99Q, as opposed to the actual trough of -7. per cent in 99Q. Second, the median counterfactual path for output growth is so much higher than the actual one that the recession would have ended (that is: GDP growth would have turned positive) in 99Q, five quarters before it actually did. By contrast, the counterfactual paths for output growth obtained by setting to zero either of the other three s are only marginally different from the actual one. Whereas the counterfactual paths for inflation are only slightly different from the actual path with the possible exception of the one obtained by killing off demand non-policy s those for the short rate exhibit two interesting results. First, absent permanent output s, the short rate would have been uniformly lower, thus reflecting a policy, on the part of Finland s monetary authority, of accomodating benign permanent supply s. Second, by the same token, eliminating demand non-policy s generates a higher counterfactual path for the short rate, reflecting the standard reaction, on the part of central banks, to the previously discussed series of negative demand s (see the the third panel of the second row of Figure 8). Finally, Figure shows how Sweden s recession of the early 99s was due to a combination of al s, with a dominant role of permanent output s, and a negligible role for demand non-policy s. The same holds true for the most recent crisis, with, again, a dominant role of permanent output s, and a negligible role played instead by monetary policy s.. Impulse-response functions to structural innovations Figures8toshowthenormalizedIRFs on impact of real GDP growth, inflation, and the short rate to each of the four identified structural s. The IRFs to a permanent output are normalized so that, for each single IRF, the long- This is conceptually in line with, e.g., the Federal Reserve s accomodation, under Alan Greenspan, of the productivity acceleration of the second half of the 99s, by keeping the Federal Finds rate comparatively low. We compute the IRFs via the Monte Carlo integration procedure described in Appendix C, which allows us to tackle the uncertainty originating from future time-variation in the VAR s structure. For reasons of space we only show IRFs on impact, but the full set of results for IRFs at horizons up to five years after the impact are available upon request.

27 run impact of log real GDP is equal to one, whereas those for the monetary policy have been normalized so that, for each quarter, the median of the posterior distribution of the impact on the short rate at zero is equal to one (which is apparent from the bottom panel of the second column of either figure). Since for neither of the other two s there exists such a natural normalization, we have normalized them by setting, for each quarter, the median of the posterior distribution of the impact at zero on the cyclical variable (either the unemployment rate, or the consumptionoutput ratio) equal to one. Once again, for reasons of space in what follows we are going to focus on the IRFs for real GDP growth, and we are going to discuss those for the other two series only when they exhibit especially interesting features. StartingfromtheUnitedStates,thekeyfeatureemergingfromFigure8isthe absence of any clear-cut peculiarity of the recent financial crisis compared to the rest of the sample period. For example, although the impact of a permanent output on real GDP growth at zero appears to have fallen compared to the years immediately preceding the crisis, such a decrease is well within the bounds of historical variation exhibited between the mid-9s and the onset of the financial crisis. The same holds true for the Euro area: although, following August 7, the impact on GDP growth of a permanent output appears to have slightly increased, the extent of variation is nowhere near that experienced during the decade of the 98s. For the Euro area, too, it is not possible to identify any obvious peculiarity of the recent crisis, compared to the rest of the sample period, for any variable and any. Turning to Japan, the crisis of the early 99s was marked by a sizeable decrease of the impact of permanent output s on GDP growth at zero, and by a corresponding increase of the impact of transitory supply s. By the same token, the reaction of the short rate to demand non-policy s exhibited a large, temporary increase around the time of the crisis, likely reflecting the more aggressive monetary policy stance adopted by the Bank of Japan in its attempt to forestall the negative headwinds generated by a series of negative demand s (see the third panel of the first row of Figure 8). This finding gains additional relevance in the light of the analogous temporary increase in the impact of such s on real GDP growth. Overall, our results paint therefore a picture in which the impact of demand s (normalized on the consumption-output ratio) temporarily increased during the crisis, thus spurring the central bank to increase, likewise, the strength of its reaction to such s. As for the recent crisis evidence is weaker, and it suggests that the impact of supply-side s, both permanent and transitory, has been qualitatively the same as during the same as, but milder than during the previous crisis. As for Finland, the crisis of the early 99s does not exhibit any clear-cut peculiar feature. The impact of demand non-policy s on real GDP growth, for example, wouldappeartohaveincreasedaroundthetimeoftecrisis,butonceagain,such an increase is well within the bounds of normal, secular variation in the rest of the sample. As for the recent crisis, the most notable features are (i) a decrease in the impact on GDP growth of permanent output s; (ii) temporary increases in the 7

28 magnitude of the impacts of transitory supply s on either variable; and (iii) sizeable temporary increases of the impact of demand non-policy s on output growth and the short rate, with peaks corresponding to the quarter of the collapse of Lehman Brothers. An important point to stress, however, is that since Finland has been, since January 999, part of the Euro area, the reaction of the short rate to structural s specific to Finland which the VAR just produces mechanically does not have a straighforward interpretation, especially in the light of the fact that Finland represents, in terms of GDP, a very small fraction of the entire Euro area, so that the European Central bank s monetary policy reacts to Finland-specific s to a very limted extent. On the other hand, it is important to keep in mind that, within the context of the recent crisis, structural s should logically be expected to have been very strongly synchronized across Euro area countries, which implies that the evidebnce reported in the bottom row of Figure is not, in fact, meaningless, and it rather difficult to interpret. Finally, as for Sweden, evidence does not point towards obviously distinctive features associated to financial crises, with the exception of a sizeable decrease in the magnitude of the impact of transitory supply s on inflation during the crisis of the early 99s, and of the impact of supply-side s, both permanent and transitory, on GDP growth during the most recent one. 5 Conclusions In this paper we have used Bayesian time-varying parameters structural VARs with stochastic volatility to investigate the specific dimensions along which recessions associated with severe financial crises have historically been different from normal recessions, and standard business-cycle fluctuations, in the post-wwii U.S., Japan, Euro area, Sweden, and Finland. We have identified four structural s by combining a single long-run restriction, to identify a permanent output as in Blanchard and Quah (989), with three sign restrictions to identify demand- and supply-side transitory s. Evidence suggests that severe financial crises have systematically been characterized by a negative impact on potential output dynamics, which in two cases Japan, following the collapse of the asset prices bubble of the second half of the 98s, and Finland, in the early 99s appears to have been nothing short of dramatic. On the other hand, with very few exceptions, neither the relative importance of the different types of s, the way they have been propagating through the economy, the persistence of output growth, or the fractions of forecast error variance explained by individual s have exhibited any systematic difference compared to standard macroeconomic fluctuations. Our main conclusion is therefore that, once controlling for the size of the s, financial crises appear to have been broadly similar to standard macroeconomic fluctuations along most, but not all, dimensions. 8

29 A The Data Here follows a detailed description of the dataset. A. United States Quarterly seasonally adjusted series for real GDP ( GDPC9: Real Gross Domestic Product, Decimal, Seasonally Adjusted Annual Rate, Billions of Chained 5 Dollars ) and the GDP deflator ( GDPCTPI: Gross Domestic Product: Chain-type Price Index, Seasonally Adjusted ) are from the U.S. Department of Commerce: Bureau of Economic Analysis. A monthly seasonally adjusted series for the civilian unemployment rate ( UNRATE: Civilian Unemployment Rate, Seasonally Adjusted, Percent ) is from the U.S. the Department of Labor: Bureau of Labor Statistics, and it has been converted to the quarterly frequency by taking averages within the quarter. A monthly seasonally adjusted series for the -month Treasury Bill ( TBMS: -Month Treasury Bill: Secondary Market Rate ) is from Board of Governors of the Federal Reserve System, and it has been converted to the quarterly frequency by taking averages within the quarter. The quarterly series for potential GDP ( GDPPOT: Real Potential Gross Domestic Product, Quarterly, Billions of Chained 5 Dollars ) is from the U.S. Congressional Budget Office, Budget and Economic Outlook. A. Euro area Quarterly, seasonally adjusted series for real GDP, the GDP deflator, and real consumption expenditure are from the European Central Bank s database. For all series, the sample period is 9Q-Q. A monthly series for a short rate, available for the period January 97-December, is from the European Central Bank s database, and it has been converted to the quarterly frequency by taking averages within the quarter. The series has been extended backwards in time to 9Q by linking it to a weighted average of short rates for nine Euro area countries, with weight set to. for Germany,. for both France and Italy, and to.5 for the remaining six countries. Since the first ten years of data (that is, 9Q to 97Q) are used to compute the (comparatively loose) priors, the fact that, up to 99Q, we rely on a reasonable proxy for the authentic, unavailable synthetic Euro area short rate is to be regarded, in our own view, as entirely legitimate. A. Japan Quarterly, seasonally adjusted series for real GDP, the GDP deflator, and real consumption expenditure are from the OECD s Quarterly National Accounts, and are available for the period 98Q-Q (their acronyms are QNA.Q.JPN.CMPGDP. VIXOBSA._S, QNA.Q.JPN.CMPGDP.DOBSA._S, and OT.Q. JPN. CMPPC. VIXOBSA.. S.VAL respectively). A monthly series for a short rate ( Official 9

30 discount rate, rate for commercial bills discounted by financial institutions for loans secured by the government ) is from the Bank for International Settlements database, and is available for the period April 955-December. A. Sweden Quarterly, seasonally adjusted series for real GDP, the GDP deflator, nominal GDP, and nominal private final consumption expenditure are from the International Monetary Fund s International Financial Statistics. The series for real GDP is available for the period 97Q-Q, whereas the other three series are available for the period period 98Q-Q. The consumption-output ratio has been computed as the ratio between nominal consumption expenditure and nominal GDP. A monthly series for a short rate ( Interest Rates, Discount Rate, Percent per Annum ) is from the International Monetary Fund s International Financial Statistics, and is available for the period January 9-December. A.5 Finland Quarterly, seasonally adjusted series for real GDP, the GDP deflator, nominal GDP, and nominal private final consumption expenditure are from the International Monetary Fund s International Financial Statistics, and are all available for the period 97Q-Q. The consumption-output ratio has been computed as the ratio between nominal consumption expenditure and nominal GDP. A monthly series for a short rate ( Interest Rates, Discount Rate, Percent per Annum ) is from the International Monetary Fund s International Financial Statistics, and is available for the period January 9-December. B Details of the Markov-Chain Monte Carlo Procedure We estimate ()-(8) via Bayesian methods. The next two subsections describe our choices for the priors, and the Markov-Chain Monte Carlo algorithm we use to simulate the posterior distribution of the hyperparameters and the states conditional on the data, while the third section discusses how we check for convergence of the Markov chain to the ergodic distribution. B. Priors For the sake of simplicity, the prior distributions for the initial values of the states and which we postulate all to be normal, are assumed to be independent both from each other, and from the distribution of the hyperparameters. In order to

31 calibrate the prior distributions for and we estimate a time-invariant version of () based on the first years of data, and we set hˆ ˆ (ˆ )i (B) where ˆ (ˆ ) is the estimated asymptotic variance of ˆ.Asfor, we proceed as follows. Let ˆΣ be the estimated covariance matrix of from the time-invariant VAR, and let be its lower-triangular Cholesky factor i.e., = ˆΣ. We set ln (ln ) (B) where is a vector collecting the logarithms of the squared elements on the diagonal of. As stressed by Cogley and Sargent (5), a variance of is huge on a natural-log scale, making this weakly informative for. Turning to the hyperparameters, we postulate independence between the parameters corresponding to the two matrices and an assumption we adopt uniquely for reasons of convenience and we make the following, standard assumptions. The matrix is postulated to follow an inverted Wishart distribution, (B) with prior degrees of freedom and scale matrix. In order to minimize the impact of the prior, thus maximizing the influence of sample information, we set equal to the minimum value allowed, the length of plus one. As for, we calibrate it as = ˆΣ, setting =., the same value used in Cogley and Sargent (5), and a slightly more conservative prior (in the sense of allowing for less random-walk drift) than the.5 used by Cogley and Sargent (5). As for, we postulate it to be normally distributed with a large variance, ( ) = (, ( ) ) (B) Finally, as for the variances of the stochastic volatility innovations, we follow Cogley and Sargent (, 5) and we postulate an inverse-gamma distribution for Var( ): µ (B5) B. Simulating the posterior distribution We simulate the posterior distribution of the hyperparameters and the states conditional on the data via the following MCMC algorithm, as found in Cogley and Sargent (5). In what follows, denotes the entire history of the vector up to time i.e. [,,, ] while isthesamplelength.

32 (a) Drawing the elements of Conditional on,, and,theobservation equation () is linear, with Gaussian innovations and a known covariance matrix. Following Carter and Kohn (), the density ( ) can be factored as Y ( )= ( ) ( + ) = (B) Conditional on and, the standard Kalman filter recursions nail down the first elementontherighthandsideof(a), ( ) = ( ),with being the precision matrix of produced by the Kalman filter. The remaining elements in the factorization can then be computed via the backward recursion algorithm found, e.g., in Kim and Nelson (), or Cogley and Sargent (5, appendix B..). Given the conditional normality of,wehave + = + + ( + ) (B7) + = + (B8) which provides, for each from - to, the remaining elements in (), ( +,,, ) = ( +, + ). Specifically, the backward recursion starts with a draw from ( ),callit Conditional on,(a7)-(a8)giveus and, thus allowing us to draw from ( ),andsoonuntil =. (b) Drawing the elements of Conditional on,,and, the orthogonalised innovations ( - ),withvar( )=, are observable. Following Cogley and Sargent (), we then sample the s by applying the univariate algorithm of Jacquier, Polson, and Rossi (99) element by element. (c) Drawing the hyperparameters Conditional on,,,and, the innovations to and to the s are observable, which allows us to draw the hyperparameters theelementsof and the from their respective distributions. (d) Drawing the elements of Finally, conditional on and the s are observable, satisfying = (B9) with the being a vector of orthogonalized residuals with known time-variying variance. Following Cogley and Sargent (5), we interpret (B9) as a system of unrelated regressions. The first equation in the system is given by, while the following equations can be expressed as transformed regressions as ³ = ³ + ³ ³ ³ = ³ ³ + For details, see Cogley and Sargent (5, Appendix B..5). (B)

33 ³ ( ) ( ) ³ = ( ) ( ) ( ) ³ ( ) ( ) ( ) ³ ( ) ( ) + where the residuals are independent standard normal. Assuming normal priors for each equation s regression coefficients the posterior is also normal, and can be computed via equations (77) of (78) in Cogley and Sargent (5, section B..). Summing up, the MCMC algorithm simulates the posterior distribution of the states and the hyperparameters, conditional on the data, by iterating on (a)-(d). In what follows, we use a burn-in period of 5, iterations to converge to the ergodic distribution, and after that we run, more iterations sampling every th draw in order to reduce the autocorrelation across draws. 5 C Computing Generalised Impulse-Response Functions Here we describe the Monte Carlo integration procedure we use in Section. in order to compute the generalised IRFs to the structural s. Randomly draw the current state of the economy at time from the Gibbs sampler s output. Given the current state of the economy, repeat the following procedure times. Draw four independent (, ) variates (the four structural s), and based on the relationship =,with [,,, ] where,,,and are the permanent output, and the interest rate, demand non-interest rate, and transitory supply structural s, respectively compute the reduced-form s at time. Simulate both the VAR s time-varying parameters and the covariance matrix of its reduced-form innovations, Ω, quarters into the future. Based on the simulated Ω, randomly draw reduced-form s from + to +. Based on the simulated, and on the sequence of reduced-form s from to +, compute simulated paths for the four endogenous variables. Call these simulated paths as ˆ +, with =,..,. Repeat the same procedure based on exactly thesamesimulatedpathsforthe VAR s time-varying parameters, the ; the same reduced-form s at times + to +; and the same structural s,,,and at time, with the only difference that, in order to compute the GIRF to,with 5 In this we follow Cogley and Sargent (5). As stressed by Cogley and Sargent (5), however, this has the drawback of increasing the variance of ensemble averages from the simulation.

34 =,,,, you set =. Call these simulated paths as =,..,. +, with For each of the iterations define + ˆ + +. Finally, compute each of the, generalised IRFs as the mean of the distribution of the + s.

35 References Benati, L. (8): The Great Moderation in the United Kingdom, Journal of Money, Credit, and Banking, 9(), 7. Benati, L., and C. Goodhart (): Monetary Policy Regimes and Economic Performance: The Historical Record, 979-8, in B. Friedman, B., and Woodford, M. (eds.), Handbook of Monetary Economics, Volume, North Holland. Benati, L., and T. Lubik (): Sales, Inventories, and Real Interest Rates: A Century of Stylized Facts, University of Bern and Federal Reserve Bank of Richmond, mimeo. Blanchard, O. J., and D. Quah (989): The Dynamic Effects of Aggregate Demand and Supply Disturbances American Economic Review, 79(), Bordo, M. D., and J. G. Haubrich (): Deep Recessions, Fast Recoveries, and Financial Crises: Evidence from the American Record, Rutgers University and Federal Reserve Bank of Cleveland, mimeo. Canova, F., and M. Paustian (): Measurement with Some Theory: Using Sign Restrictions to Evaluate Business Cycle Models, Journal of Monetary Economics, 58, 5. Carter, C. K., and R. P. Kohn (): On Gibbs Sampling for State Space Models Biometrika, 8, Cerra, V., and S. C. Saxena (8): Growth Dynamics: The Myth of Economic Recovery, American Economic Review, 98(), Cochrane, J. H. (99): Permanent and Transitory Components of GNP and Stock Prices, Quarterly Journal of Economics, 9(), 5. Cogley, T. (5): How Fast Can the New Economy Grow?, Journal of Macroeconomics, 7(), Cogley, T., and T. J. Sargent (): Evolving Post-WWII U.S. Inflation Dynamics, in B. Bernanke and K. Rogoff, eds. (), NBER Macroeconomics Annuals. (5): Drifts and Volatilities: Monetary Policies and Outcomes in the Post WWII U.S., Review of Economic Dynamics, 8(April),. Cogley, T. W. (99): International Evidence on the Size of the Random-Walk in GNP, Journal of Political Economy. 5

36 DelNegro, M. (): Discussion of Cogley and SargentŠs Drifts and Volatilities: Monetary Policy and Outcomes in the Post WWII U.S., Federal Reserve Bank of Atlanta, Working Paper -. Fry, R., and A. Pagan (7): Some Issues in Using Sign Restrictions for Identifying Structural VARs, NCER Working Paper n., April 7. Gali, J., and L. Gambetti (9): On the Sources of the Great Moderation, American Economic Journal: Macroeconomics, (), 57. Gali, J., F. Smets, and R. Wouters (): Slow Recoveries: A Structural Investigation, Journal of Money, Credit, and Banking, forthcoming. Gambetti, L., E. Pappa, and F. Canova (): The Structural Dynamics of U.S. Output and Inflation: What Explains the Changes?, Journal of Money, Credit, and Banking, (-), Golub, G., and C. VanLoan (99): Matrix Computations, III Edition. Johns Hopkins University Press, Baltimore. Jacquier, E., N. G. Polson, and P. Rossi (99): Bayesian Analysis of Stochastic Volatility Models Journal of Business and Economic Statistic,, 7 8. Kim, C. J., and C. Nelson (): State-Space Models with Regime Switching. Cambridge, Mass., The MIT Press. Lopez-Salido, D., and E. Nelson (): Postwar Financial Crises and Economic Recoveries in the United States, Federal Reserve Board, mimeo. Orphanides, A. (): Monetary Policy Rules Based on Real-Time Data, American Economic Review, 9(), Orphanides, A. (): Monetary Policy Rules and the Great Inflation, American Economic Review, 9(May ), 5. Orphanides, A. (a): Historical Monetary Policy Analysis and the Taylor Rule, Journal of Monetary Economics, 5, 98. (b): The Quest for Prosperity Without Inflation, Journal of Monetary Economics, 5(April ),. Papell, D. H., and R. Prodan (): The Statistical Behaviour of GDP Following Financial Crises and Severe Recessions, University of Houston, mimeo. Primiceri, G. E. (5): Time Varying Structural Vector Autoregressions and Monetary Policy, TheReviewofEconomicStudies, 7, 8 85.

37 Reinhart, C., and K. Rogoff (9): This Time Is Different. Princeton University Press. Rubio-Ramirez, J., D. Waggoner, and T. Zha (5): Structural Vector Autoregressions: Theory of Identification and Algorithms for Inference, Review of Economic Studies, 77(), 5Ů 9. Stock, J. H., and M. W. Watson (): Disentangling the Channels of the 7-9 Recession, Brookings Papers on Economic Activity, forthcoming. Taylor, J. B. (9): Getting Off Track. Hoover Institution Press. 7

38 Figure United States, Euro area, and Japan: estimated transitory components of log real GDP

39 Figure United States, Euro area, and Japan: log real GDP and estimated permanent component

40 Log real GDP, and log permanent real GDP Transitory component of log real GDP Finland Beginning of EMU Collapse of Lehman Brothers. Introduction.7. of inflation targeting. Median. Sweden.5. 8th percentile th percentile Log real GDP Beginning of the financial crisis Figure Finland and Sweden: log real GDP, and estimated permanent and transitory components

41 Figure Financial crises and potential output dynamics: posterior distribution of the difference between the annualized peak-to-trough change in potential output during the crisis, and its annualized change during the previous five years (for Japan s early 99s crisis, the comparison is between the periods before and after the crisis)

42 Figure 5 Standard deviation of reduced-form innovations to the log-difference of real GDP, and log-determinant of Ω t

43 Figure Output growth persistence

44 Figure 7 United States and Euro area: fractions of draws for which the structural s have been positive 5

45 Figure 8 Japan and Finland: fractions of draws for which the structural s have been positive

46 Figure 9 Sweden: fractions of draws for which the structural s have been positive 7

47 Permanent supply Monetary policy Demand non-policy Transitory supply 5 Real GDP growth Inflation Actual.5 series Short rate Collapse of Lehman Brothers 8th percentile Median th percentile Figure United States, counterfactuals for the Great Recession shutting off one structural at a time

48 Real GDP growth Inflation Short rate Permanent supply Actual series Monetary policy Demand non-policy Transitory supply th percentile 8 9 Collapse of Lehman Brothers 8th percentile Median 8 9 Figure Euro area, counterfactuals for the Great Recession shutting off one structural at a time

49 Permanent supply Monetary policy Demand non-policy Transitory supply 8 Real GDP growth Inflation Short rate th percentile 8th percentile Median Actual series Figure Japan, counterfactuals for the financial crisis of the early 99s shutting off one structural at a time

50 Permanent supply Monetary policy Demand non-policy Transitory supply 5-5 Real GDP growth Inflation Short rate th percentile.5 Actual series Median th percentile 8 9 Figure Japan, counterfactuals for the Great Recession shutting off one structural at a time

51 Permanent supply Monetary policy Demand non-policy Transitory supply Inflation Short rate 8 9 Actual 5 8th 8 series percentile th percentile Median Real GDP growth Figure Finland, counterfactuals for the financial crisis of the early 99s shutting off one structural at a time 5

52 Real GDP growth Inflation Short rate Permanent supply 5-5 Collapse of Lehman Brothers Monetary policy Demand non-policy Transitory supply th percentile Median th percentile -5 Actual - series Figure 5 Finland, counterfactuals for the Great Recession shutting off one structural at a time

53 Permanent supply Monetary policy Demand non-policy Transitory supply 8 Inflation Real GDP growth Short rate th Median percentile Actual series 8th percentile Figure Sweden, counterfactuals for the financial crisis of the early 99s shutting off one structural at a time 7

54 5-5 Real GDP growth Inflation 5 5 Short rate 5 8th Permanent percentile supply Median -5 th Actual series percentile Monetary policy Collapse of Lehman Brothers Demand non-policy Transitory supply Figure 7 Sweden, counterfactuals for the Great Recession shutting off one structural at a time 8

55 Figure 8 United States, impulse-response functions to structural s on impact

56 Figure 9 Euro area, impulse-response functions to structural s on impact 9

57 Figure Japan, impulse-response functions to structural s on impact

58 Figure Finland, impulse-response functions to structural s on impact

59 Figure Sweden, impulse-response functions to structural s on impact