In recent writings, Carmen M. Reinhart and Kenneth S. Rogoff have drawn lessons from previous financial crises to predict the fallout from the current crisis (see Reinhart 2009 and Reinhart and Rogoff 2008a,b, 2009). We do have sympathy for their conclusion, but would like to sound a note of caution. As we argue in our contribution to last August’s Jackson Hole Symposium (Cecchetti et al. 2009), while they may be similar in their sources, financial crises appear quite different with respect to their ultimate impact. Any attempt to learn from the past must, in our view, take these differences into account.

We analyse 40 systemic banking crises in 35 countries and find substantial variation. Some were associated with no downturn whatsoever (Bolivia 1994, Colombia 1982, Nicaragua 2000, Paraguay 1995, Sri Lanka 1989 and Vietnam 1997), while several others were associated with contractions similar in magnitude to those arising from ordinary recessions. At the extreme, a small number of crises were both extraordinarily protracted and very deep. For example, Bulgarian real GDP fell 42% in the mid-1990s, and it took almost seven years for the country to recover its pre-crisis level of output (although factors other than the financial crisis clearly played an important role in that episode).

Crises appear even more heterogeneous when we consider dimensions beyond their depth and length. We base this conclusion on cluster analysis, a method that assigns observations (crises) to subsets that are similar, given a set of characteristics. Cluster analysis is widely used in quantitative social research to analyse datasets encompassing a large number of variables.

In our work, we consider 28 financial crises that have occurred since the 1980s. For each crisis, we collected information on country and crisis characteristics, whether or not there had been a boom in the run-up to the crisis, macroeconomic vulnerabilities, the policy response during the crisis, and external conditions in the years after the start of the crisis.

The results of the cluster analysis are represented in a dendrogram (Figure 1). It shows the Euclidian distance (a measure of dissimilarity with respect to all variables) between the crises along the horizontal axis; the country and year of the crisis are on the vertical axis. If the Euclidian distance between two observations is below a given threshold level on the horizontal axis – that is, they are more similar than the level of dissimilarity we allow – they are joined in a cluster. Observations with distances above the threshold remain separate. Thus, in general, each observation would form its own cluster if the threshold distance is set at zero, and all observations would fall into one cluster if the threshold is set to be sufficiently large.

Figure 1. Cluster analysis (Current crises are capitalised)


This time is different

One of the most striking conclusions we draw from this way of looking at the data is that current events are unique. Indeed, the current crisis is less similar to all of the crises in our database than, say, the Japanese financial crisis of the 1990s is to either the crisis experienced by Ecuador in 1998 or the one that took place in Bulgaria during the transition! From this we conclude that it is not possible to learn about the likely path of the current crisis by averaging outcomes across past crises.

At first sight, this result may seem discouraging, but it is not. After all, variation is the econometrician’s friend. With that in mind, we look to see how particular characteristics of either crises or the environment in which they take place affect crisis outcomes.

Our findings suggest that the contractions associated with systemic banking crises are longer and deeper when they are accompanied by a currency crisis or when growth is low immediately prior to the onset of the crisis. And we find that a systemic banking crisis is less costly when it is accompanied by a sovereign debt default. However, this last finding is almost surely driven by the fact that such debt is usually held mainly by foreigners, so in these cases a portion of the cost of the crisis is being exported to non-residents.

More intriguing than what we find is what we expected to find but could not. For example, there is no evidence that either the income level of a country, as measured by per capita GDP, or external conditions, such as growth in the largest trading partners or the number of concurrent crises, affect the length or depth of a crisis-related contraction. This last point is particularly interesting since, if taken at face value, it means that the global nature of the current crisis should not slow down recovery relative to the benchmark of a crisis in a single country. That said, we will have to see whether this conclusion stands the test of time.

Our estimates allow us to predict how long we expect it to take for output in a country currently experiencing a crisis to recover to its pre-crisis level. Taken at face value, they suggest that the major economies currently affected will regain their pre-crisis levels of output by the second half of 2010 (Figure 2) – that is, roughly one year from now – although the confidence interval around this prediction is very large. In some cases, this may seem optimistic as it implies relatively high growth rates in the near term. While we continue to maintain that there is substantial variation across crises, it is fairly common to see rather rapid recoveries.

Figure 2. Predicted real output costs of the current crisis for selected countries
Length of contraction, in quarters

Depth, as a percentage of GDP at peak level Cumulative output loss, as a percentage of GDP at peak level

CH = Switzerland; DE = Germany; ES = Spain; GB = UK; IE = Ireland; JP = Japan; NL = Netherlands; US = United States. Predictions for current crisis using preferred model and preferred model including depth; with 90% confidence intervals.

Regardless of whether crisis-country output returns to its pre-crisis level slowly or quickly, it is still likely to have lasting costs.

  • First, there are the missed years of growth that would presumably have happened in the absence of the crisis.
  • Second, there is the very real possibility that output growth will be permanently slower as a consequence of the crisis (although, admittedly, the real estate boom in the US and other economies in the run-up to the crisis may also have boosted output growth).

Our results are consistent with this common view, as we find that many systemic banking crises have had a long-lasting negative impact on the level of GDP. And even in those cases in which trend growth was higher after the crisis than it had been before, making up for the output loss resulting from the crisis itself took years.


Cecchetti, S G, M Kohler and C Upper (2009), “Financial crises and economic activity”, paper presented at the symposium on Financial stability and macroeconomic policy organised by the Federal Reserve Bank of Kansas City, Jackson Hole, 20–21 August.

Reinhart, C M (2009), “The economic and fiscal consequences of financial crises”,, 26 January.

Reinhart, C M and K S Rogoff (2008a), “Banking crises: an equal opportunity menace”, NBER Working Paper 14587.

Reinhart, C M and K S Rogoff (2008b), “The aftermath of financial crises”, NBER Working Paper 14656.

Reinhart, C M and K S Rogoff (2009), This time is different: eight centuries of financial folly, Princeton University Press.

420 Reads