The Length and Cost of Banking Crises

Contributor Notes

Author’s E-Mail Address:

This paper reviews how recent studies of banking crises differ with regard to the dating, length, and costs of the crises. Significant discrepancies in these features suggest the absence of analytical consensus. The data allow an examination of the relation between perceived crisis length, as an index of delay in taking actions to resolve a crisis, and crisis costs. Cross-sectional evidence does not show that the length of a crisis is a significant contributor to its resolution cost. A measure of economic cost, the growth shortfall in the crisis period, shows more evidence of a link.


This paper reviews how recent studies of banking crises differ with regard to the dating, length, and costs of the crises. Significant discrepancies in these features suggest the absence of analytical consensus. The data allow an examination of the relation between perceived crisis length, as an index of delay in taking actions to resolve a crisis, and crisis costs. Cross-sectional evidence does not show that the length of a crisis is a significant contributor to its resolution cost. A measure of economic cost, the growth shortfall in the crisis period, shows more evidence of a link.


The spread of banking crises globally in recent decades has spurred interest in studying the phenomenon. One of the principal problems of investigating banking crises is identifying when such occurrences happened and how long they lasted. Accurately timing banking crises is obviously important to the financial surveillance responsibilities of the Fund in terms of determining the best policies for preventing their onset and limiting their damage. A banking crisis is not a simple event that takes place at a point in time, like a revaluation of a currency parity under the Bretton Woods system, for example. Rather, it is an episode, spread out diffusely over time, with no clear beginning or end.

An “objective” procedure cannot readily be used to date episodes of banking crises as has been done in the case of currency crises. Some studies of currency crises identify them by unusually large declines in a weighted index of exchange rates and international reserves (and possibly interest rates).1 Bank loans by their nature do not allow high-frequency observations of their underlying market value that can be used to construct an empirically-based indicator for dating crisis episodes. Also, critical distress in a country’s banking system can show up in many ways. It may not always manifest itself as a sharp decline in the value of bank assets or in the size of the deposit base. It may result instead in special official assistance, nationalization, forced mergers, or a variety of changes in circumstances that cannot be fully anticipated by a simple quantitative index.

Nor will every instance of distress or weakness in a country’s banking system represent a crisis. Even a large individual bank may fail because of fraud or mismanagement. A group of banks may fail because of undiversified exposure to regional economic risks. If such problems are limited, and perceived as resulting from shocks that are specific to one institution or a narrow class of institutions, they need not pose a risk of systemic crisis.

As a consequence, classifying and dating a banking crisis involves some irreducible degree of subjective judgment. In this respect, the problem of dating a banking crisis is similar to that of designating a recession. Recessions are treated as recurrent, although not periodic, events that have a fundamental similarity in terms of a sustained, general decline in economic activity followed by a sustained, general recovery. This fundamental repeating pattern of macroeconomic dynamics makes recessions identifiable events that are subject to analysis. Although simple rules have been proposed to define recessions—for example, two or more quarters of real output decline—the consensus of economists views recessions as too complicated to be so simply characterized. The breadth of output declines, the conditions of the labor market, and many other factors are also important. So economists have come to rely on expert judgment to demarcate the timing of recessions. In the United States, for example, the National Bureau of Economic Research has acquired that role by virtue of its traditional expertise on business cycle indicators.

Banking crises are events similar to recessions. Crises share a pattern of widespread damage to banking systems; like recessions, their timing cannot be easily determined by simple statistical filters, such as the ratio of nonperforming loans, and is better established by careful expert judgment. As yet, however, no consensus has been achieved about who should be the arbiter of dating banking crises.

Appropriately dating banking crises and determining their length and cost is important for assessing how much delay in dealing with a crisis may add to its cost. Data from these recent studies allow a direct examination of the relation between bank crisis length and measures of crisis cost.

The paper is structured as follows. Section II reviews recent studies of banking crises and examines how they differ regarding the dating, length and costs of crises. Section III uses data from these studies to analyze the relationship between crisis length and resolution costs. Even after incorporating various determinants of initial crisis costs, a significant relation between length and cost cannot be established. Section IV considers the relationship between crisis length and a more general measure of economic costs, the shortfall in economic growth during the crisis period. Here the evidence is somewhat more supportive of a positive relation between length and cost, but the connection is again not very strong. Section V presents conclusions.

II. Recent Studies

The approach of most researchers in this area has been not to impose wholly their own judgment, but to incorporate the judgments of observers with expertise about country banking systems. This paper reviews five recent studies of banking crises.2 Appendix 1 summarizes the episodes listed in each study and provides information, when available, on the dating and severity of the episodes.

The criteria used in the studies for the selection and dating of banking crises are summarized below:

  • Caprio and Klingebiel (CK). This study organizes data on episodes of bank insolvency since the late 1970s. “It relies upon the assessment of a variety of finance professionals. Only published sources or interviews with experts familiar with individual episodes were employed. The dates attached to the crises reviewed here are those generally accepted by finance experts familiar with the countries. Some judgment has gone into this list [of systemic episodes]…virtually every transitional economy [TE] at some stage in the transition process belongs on the (systemic) list; however, in the interest of limiting the number of countries with missing information these were excluded. Including all the TEs would bring the number of countries covered to about 90 and the episodes to well over 100.”

  • Demirguc-Kunt and Detragiache (DKD). “…we have identified and dated episodes of banking sector distress during the period 1980-94 using primarily five recent studies: Caprio and Klingebiel (1996), Drees and Pazarbasioglu (1995), Kaminsky and Reinhart (1996), Lindgren, Garcia and Saal (1996), and Sheng (1995). Taken together, these studies form a comprehensive survey of banking sector fragility around the world. We established—somewhat arbitrarily—that for an episode of distress to be classified as a full-fledged crisis in our panel at least one of the following four conditions had to hold:

    • (i) The ratio of nonperforming assets to total assets in the banking system exceeded 10 percent.

    • (ii) The cost of the rescue operation was at least 2 percent of GDP.

    • (iii) Banking sector problems resulted in a large scale nationalization of banks.

    • Extensive bank runs took place or emergency measure such as deposit freezes, prolonged bank holidays, or generalized deposit gurantees were enacted by the government in response to the crisis.

In a new cases, however, we had insufficient information and made a decision based on our judgement.”

  • Dziobek and Pazarbasioglu (DP). “The sample consists of a representative group of countries reflecting a broad coverage across regions and levels of development. Countries were included only in cases where the problems were judged to be systemic…defined as a situation where problems affected banks which, in aggregate, held at least 20 percent of the total deposits of the banking system. A questionnaire was sent to country authorities and, in some cases, to IMF or World Bank staff with special expertise on banking sector problems. Data were requested for three points in time; the year when bank restructuring action started four years before and four years after that date…This information and data covered five broad areas: banking structure; bank performance; banking sector institutional framework (regulatory, legal and accounting environment); instruments of bank restructuring used; and, costs and budgetary implications.”

  • Kaminsky and Reinhart (KR). “. we mark the beginning of a banking crisis by an event that indicates either: (i) Bank runs that lead to the closure, merging, or takeover by the public sector of one or more financial institutions…(ii) If there are no runs, the closure, merging, takeover, or large-scale government assistance of an important financial institution (or group of institutions), that marks the start of a string of similar outcomes for other financial restitutions [sic]…To date these events, we rely heavily on existing studies of banking crises and on the financial press around the time of the crisis.”

  • Lindgren, Garcia, and Saal (LGS). “Sources…include IMF desk economists, Sheng (1996), Caprio and Klingebiel (1996), the studies in Sundararajan and Balino (1991), and various official and news publications. Two general classes are identified: ‘crisis’… and ‘significant’ problems… There is some degree of judgment in these classifications, but in general we refer to cases where there were runs or other substantial portfolio shifts, collapses of financial firms, or massive government intervention, as crises. Extensive unsoundness short of a crisis is termed significant.”

CK, DKD and LGS generally specify both starting and ending dates at annual frequency for crises, although in several cases crises that have emerged recently are treated as ongoing. DP specify only the year in which bank restructuring operations began, which is an objective, observable criterion but which does not really date the emergence of the crisis. KR specify only the start of the crisis but do so at a monthly frequency.3

The classification of episodes of banking crises in these studies relies fundamentally on the judgments of expert opinion and more derivatively on the work of earlier studies. In this regard, CK seem to be a root source, since they base their classification completely on surveys of expert opinion and they are used as a source by several other studies. None of the studies suggest that the timing set by these subjective judgments was adjusted by using more objective information sources, such as changes in risk premia for sovereign or domestic bank borrowers in international markets, overall country risk ratings, the volume of banking system deposits, etc. Reliance on subjective expert views to date crises is appropriate. Simple multiple objective indicators, whether based on bank balance sheet data or financial market prices, are likely to be misleading. The concept of a crisis is meaningful principally in terms of its effects on the expectations of participants in financial markets and the economy. Expert opinion can directly reflect these expectations. The risk of error arises from expert opinion not being representative of the broader range of expectations in markets and the economy.

There are discrepancies among the various studies regarding the timing and costs of banking crises. Table 1 shows that LGS are the most comprehensive listing of banking crisis episodes, with 149 occurrences. CK list a total of 86 episodes but has deliberately excluded several potential instances, especially among the transition economies, in order to limit the number of occurrences for which useful data are missing. CK and LGS have a much smaller discrepancy in the number of crisis episodes that have a definite length. For many episodes, in both studies, crises are listed as taking place in indefinite periods, such as “the 1980s” or as being “ongoing” or having a definite starting date but remaining open-ended. LGS use stricter criteria in practice for the selection of systemic banking crises than does CK; LGS identify 40 systemic episodes to 72 for CK.

Table 1.

Summary Comparison of Various Indicators

article image

The other studies have a smaller set of episodes than CK and LGS because they were constructed to analyze the determinants or consequences of banking crises—rather than to provide comprehensive lists—and were generally limited by the availability of other information needed for the analysis. In some cases the selection procedures have produced a biased sample. As Appendix 1 shows, both DKD and KR exclude observations from transition economies and KR also omit African episodes. The exclusion of only the TEs may not create a severe selection bias. Banking crises in those countries can reasonably be viewed as the result of historically unique stresses. They are not a large fraction of total observed crises: TEs account for 25 of the total 149 episodes in the comprehensive LGS list and only five of the 40 systemic episodes.4 However, the omission of both Africa and the TEs from the KR sample may pose a more serious problem for generalizing any conclusions about banking crises from that study. Africa and the TEs together account for 75 of the 149 total LGS episodes and 19 of the 40 systemic cases. The DP set avoids this problem because it was constructed to be a geographically representative sample.

A comparison of the two comprehensive sets—CK and LGS—reveals significant differences in the timing of common episodes. Table 2 shows comparative data for two subsets of common observations. There are 35 common observations between the two studies for crisis episodes of definite length (CK show 67 instances; LGS, 79). Of these 35, only nine show identical crisis periods.

Table 2.

Comparison of CK and LGS Observations on Banking Crises

article image

Often, however, the feature of interest is the starting date rather than the entire period of the crisis. Of the 35 common episodes, 20 have the same starting date and five others differ by only one year. This way of looking at the episodes shows a closer correspondence between the two studies. Still, more than 40 percent of the common episodes differ by a year or more in their start. If economic observers had similar disagreements about the timing of recessions, the concept of recession would be seriously impaired as a tool of empirical research. If the comparison is limited to systemic episodes with a definite starting date, overlap improves somewhat. Of 26 such crises that the two studies have in common, 17 have the same starting year and an additional four differ by one year.

Unless different observers share a common perception of what the facts are about banking crises around the world, developing any understanding about the process will be fundamentally hampered. Dating these crises is not the only dimension where perceptions differ. Table 3 provides data on the length and resolution costs of 11 overlapping episodes for which CK and LGS have complete data for both features. Of the 11 instances, four have an exact overlap in their periods, eight share the same starting date and two others have a one-year difference in the starting date. This is a close agreement on starting dates, but the studies differ somewhat on the average length of crisis. Furthermore, there is a more pronounced difference regarding the size of resolution costs, although much of this is attributable to the discrepancy for Argentina.5

Table 3.

Comparison of CK and LGS Observations, Common Episodes of Definite Length and Observed Resolution Costs

article image

III. The Relation Between Crisis Length and Resolution Cost

The data sets on global banking crises allow for a cross-section exploration of what effect delay in resolving a crisis has on the cost of the crisis. Most analysts of banking crises argue for their prompt resolution because of the risks posed by moral hazard. Banks that have negative or low net worth, and operate under limited liability with explicit or implicit depositor protections, have incentives to gamble. Such banks are willing to undertake investments with a low probability of a high payoff but which do not cover their risk and, therefore, pose an expected loss. As every croupier knows, anyone who keeps making unfair bets eventually loses. The longer that “dead” banks gamble with public insurance, the larger will be the public sector bill.

This tendency has a real cost because resources are directed to uneconomic uses and the creditworthiness of the government winds up encumbered, reducing its ability to give financial support to socially useful purposes. This theoretical argument is very general and should apply to varied circumstances since it depends only on the financial conditions of banks operating in a nearly universal legal and regulatory environment. As Caprio and Klingebiel (1996a) put it: “The common link between different types of bank failure is that the initial losses, whatever their cause, often multiply when prompt corrective action is not taken.” Or Lindgren, Garcia, and Saal: “Delay in addressing unsound banks is rarely effective and usually detrimental.... unsound banks tend to take on even more risk or may be looted by insiders, which ultimately increases resolution costs.”

A natural assumption is that delay in taking effective corrective action is directly related to the length of a crisis. Crisis length, then, can be used as a proxy measure for delays, and differences in crisis length across countries can measure differences in the speed of acting to effectively resolve the problem. Episodic evidence from different cases suggests the costliness of delaying resolution. However, systematic evidence across countries of a positive relation between the length of a banking crisis and the amount of resolution costs is not abundant.

DP show a negative relation between the time delay in taking bank restructuring measures after the surfacing of systemic problems and an overall measure of success in dealing with those problems. Their results show a general relation between rapid action and more progress in addressing bank problems but do not give indications of the strength or significance of the relationship. The success measure does not directly contain resolution costs but uses a number of factors that are clearly determinants of those costs—such as nonperforming loans and profitability indicators—as well as indicators of overall banking system performance. And the measure of the length that they consider refers to the gap between the onset of systemic problems and restructuring actions. This gap is not equivalent to the full length of a banking crisis, which may begin before systemic problems are perceived (some crises are not classified as systemic) and will extend beyond the onset of restructuring actions until some degree of stabilization is achieved in the banking system.

The CK and LGS data sets, however, do allow a direct test of the correlation between the length of a banking crisis and its costs. It should be noted at the outset that these cost data are not comprehensive or fully reliable. The costs reported in these data sets are meant to be the fiscal costs associated with the resolution of the banking crisis. Their accuracy varies across countries, however. In some cases they are based on scanty information; in others, a broader measure than just fiscal costs may have been inadvertently incorporated. The resolution costs of relatively recent crises are subject to constant revision as new losses or stronger-than-anticipated recoveries materialize. They do not include, as Caprio and Klingebiel (1996b) write, “that part of the burden born by depositors and borrowers in the face of widened spreads for bad loans that were left on banks’ balance sheets” or the costs of indirect assistance such as “giving a subsidy to a borrower, granting the borrower some monopoly privilege or other means to improve its profits and thereby repay loans, or by directly injecting funds to banks.” Also omitted are the macroeconomic costs that arise through the disruption to the credit system from bank failures or in other ways, which may be substantial since bank crises are frequently associated with succeeding recessions. Finally, the implicit assumption that direct resolution costs are correlated with total costs may be wrong in some instances. A country, for example, could avoid a banking crisis by resorting to massive monetary expansion, thereby passing the burden onto the economy at large.

Crisis costs, of course, depend on factors other than just the delay in addressing banking system problems. They depend foremost on the nature and magnitude of the shocks that initiated the crisis – a general recession, a terms-of-trade decline, the collapse of a bubble in the property market, and so forth. They will depend as well on the types of actions that the authorities take to resolve the crisis. Some actions may be less costly than others; for example, a merger of a failed institution into a viable one on assisted terms may be less costly than closure and liquidation because it avoids the “fire sale” disposition of relatively illiquid assets. If the size of the initial crisis shocks and the types of resolution actions undertaken are independent of the length of delay in taking action, no bias will be introduced into an ordinary least squares (OLS) cross-section estimation of the relation between crisis length and cost. For the initial estimations, such independence is assumed. The need to re-examine the assumption of independence between crisis length and the size of the shocks that trigger a crisis will be discussed later. However, the assumption of independence between crisis length and the types of resolution actions undertaken is broadly justifiable.

Assume that when the authorities take action, they will choose the option that is lowest cost. Then, for assessing the relationship between delay and crisis cost, what matters is whether the menu of options available for actions to resolve a banking crisis varies systematically with the length of the delay in taking action. Specifically, the estimated relation between cost and delay will be biased downward if waiting longer to act allows more choices for dealing with the crisis to emerge. A wider number of policy options will be likely to increase the prospects of undertaking lower cost resolution actions. If a longer time means more options, and more options mean lower costs, a longer-lived crisis may be less costly, even given the moral hazard incentives that delay creates for increased crisis costs.

Does waiting to take resolution actions systematically increase of decrease the policy options that authorities have? This answer is not clear. Consider first a comparison of closure versus liquidation. The authorities always have the option to close and liquidate a failed bank, but this action may be a more costly resolution than merger with a viable institution. Developing a merger option, however, may take time. In this case, waiting to act may reduce resolution costs. On the other hand, suppose a merger option exists. Delay in acting, say, in the hope of developing better merger alternatives, risks losing the original option, for example, through unfavorable changes in the financial condition of the likely acquirer. The absence of a clear systematic relation between crisis length and the menu of options for resolution actions means that omitting the types of policy actions taken does not bias the estimation of the relation between crisis length and cost.

The CK data set includes a sample of 29 countries that have a banking crisis with both the definite length of the crisis period and the resolution costs (as a percentage of GDP) identified. These are listed in Table 4.

Table 4.

Length and Cost of Banking Crises (CK Sample)

article image

Figure 1 shows the scatter diagram for these data with a trend line and Equation (1) reports the results of a simple OLS regression of cost on length.

Figure 1:
Figure 1:

Cost of Banking Crises vs Length, CK Sample

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

* significant at 1 percent confidence level
Figure 2:
Figure 2:

Cost of Banking Crises vs. Length, LGS Sample

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

There is no significant relationship between the length of a banking crisis and its resolution costs in the CK data. Although the coefficient on length is positive, it is far from significant and the extremely low R2 indicates an absence of explanatory power.6

Similar results occur using LGS data for episodes of definite length where information on resolution costs is given. Table 5 provides these data and Equation (2) shows again another insignificant positive relation with no explanatory power.

Table 5.

Length and Cost of Banking Crises (LGS Sample)

article image

There is not a clear-cut relation between crisis length and cost in the cross-sectional data despite a potential bias in that direction. The observed length of a crisis period may not measure the conceptually appropriate delay between the onset of a problem and actions to resolve it effectively. Episode lengths reported in the CK and LGS samples reflect experts’ judgments about the beginning of the crisis. There is, however, a necessary perception lag between first impressions of a crisis and the time at which banks actually built up large losses. Observers may have an ex-post tendency to shorten this perception lag by recalling their perceptions of a crisis to be earlier in the process than they actually were. This tendency would be especially strong for major crises: their standing as experts is enhanced more, the sooner they are perceived to have called the onset of a major crisis.7 Unless dating is done against a contemporaneous record of expert perceptions (and this does not seem to be the case throughout the CK and LGS samples), the temptation to enhance reputation ex post will work to bias upward the perceived length of costly bank crises.

Why is there not a significant link?

Various reasons may account for the somewhat surprising lack of a significant positive correlation between crisis length and resolution costs.

The reduced borrowing capacity of low net worth banks

One factor that can produce a weak linkage between crisis length and cost is the negative effect of asset losses on the ability of troubled banks to raise funds. This effect can arise even in a regime of implicit deposit insurance. Implicit deposit insurance is expected to protect the capital value of bank deposits but provides no assurance about their liquidity. So losses that threaten to put a bank into a negative net worth position raise the probability that depositors will exercise their (implicit) call on the deposit insurance fund. This risk, in turn, reduces the liquidity of any new deposits with the institution and works to tighten counterparty limits and restrict new funds — a kind of residual market discipline.

Additionally, if enough banks fall to a negative net worth position during a crisis, the amount of unencumbered reserves that support the deposit insurance function may be severely reduced. These reserves would have to be reconstituted through a government appropriation and there will be some uncertainty whether that replenishment would be done in full, depending on the government’s own fiscal position. As a consequence, the marginal attractiveness of new deposits for all banks in a system undergoing crisis will be reduced.8

Under these conditions, low net worth banks will find their access to new deposits restricted. Although they have an incentive to gamble by acquiring unjustifiably risky assets, their ability to take on such assets quickly is limited. Since the growth in the total footings of low net worth banks is checked by their restricted capacity to attract new funding, they can only take on new risky assets by substituting for old, less risky ones.9 This slow process limits the rate at which new risk is undertaken. As a result, banks in crisis do not rapidly generate new losses and add to crisis resolution costs.

Subtle intervention

The manner in which supervisory intervention is conducted may also affect the perceived relation between banking crisis length and resolution cost. Removing the distorted incentives for gambling facing the management of banks that have suffered major net worth erosions is achieved usually by overt actions such as directed recapitalization, forced shrinkage, replacement of management, merger, etc. These overt actions are typically the events that signal to observers the denouement of a crisis episode. It is possible, however, for the authorities to influence behavior through subtle interventions that may not be readily apparent to outside observers. Discreet warnings to a troubled bank to adopt a more conservative business posture may result in that bank going into standstill mode even without a change in management. In that case, no new unjustifiable risks will be undertaken. Even if the condition of damaged assets worsens further, such losses may be covered by earnings on viable assets. That bank, then, will appear to an outside observer to be in an unresolved crisis. It will have low net worth, low earnings, and no overt actions will have been taken to address the matter. But at the margin, behavior will have moved away from dangerous risk-taking. These conditions will be viewed as an ongoing, long-lived crisis that does not generate further increases in resolution costs, contrary to presumptions.

Faster reaction to larger crises

A final important factor interfering with a strong positive link between crisis length and cost is an incentive for authorities to act more quickly when the initial perception of the size of the crisis is higher. This tendency undercuts the assumption of independence between initial shocks and crisis length mentioned earlier. The process of a banking crisis can be conceptually divided into different stages (Figure 3). The first stage is the period between the onset of crisis problems and the first awareness of them outside the problem banks; call this the perception lag. The second stage, the action lag, is the gap between those initial perceptions and the perceived resolution of the crisis through policy measures. The action lag can be further divided into “inside” and “outside” action lags. The inside action lag is the gap between the perceived beginning of the crisis and the initial policy steps to resolve it. The outside lag is the period from those initial actions until the actions are perceived to be effective in ending the crisis. This last lag may be prolonged if the initial policy actions were insufficient to resolve the crisis and must be followed by another set of measures. Such a sequence of initial failure and repeated follow-up may account for occasional differences among observers about whether a country has experienced one long crisis or two (or more) separate episodes.10 Conversely, if the initial policy actions are generally viewed as very strong, the outside action lag may be squeezed down to zero length. The observed crisis length will run from the end of the perception lag to the end of the outside action lag.

Figure 3:
Figure 3:

Temporal Sequence of Crisis Resolution

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

The perception lag may vary in length depending on the nature of the shocks that trigger problems and the way that banks respond. Initial disturbances—such as exchange rate depreciation, terms of trade declines, falls in property values, and so forth—can obviously vary in their intensity. Troubled assets may be disguised by “evergreening” of loans through a complicated network of related borrowers, by off-balance sheet manipulations, or by outright fraud. As a consequence, initial perceptions of the size of a banking sector problem and its implied resolution costs can vary a great deal from case to case.

If the first impressions of the size of a problem are large, the authorities may be very prompt in their response. The prudent assumption for bank regulators is to assume that conditions are deteriorating at a rate commensurate with the initial size of the problem. This assumption is consistent with the incentives facing the banking system. The larger the initial losses arising from the original shocks, the greater is the likely amount of assets in banks with low net worth. The larger, in turn, is the likely amount of loss-provoking gambling that will be undertaken and the faster will the crisis worsen. The authorities, then, have a greater incentive to act quickly to resolve an initially large crisis because the incentives facing the banking sector will make the situation deteriorate rapidly. On the other hand, incentives to address a smaller initial problem are weaker, since the prospective rate of deterioration is less.

So crises that show a high perceived cost at the end of the perception lag will have a short action lag and, probably (if the initial actions are vigorous) a short crisis length. Episodes that show a relatively low cost at the end of the perception lag will not spur such quick action. If banking sector conditions in those cases worsen relatively slowly, they will result in a lengthy crisis but with relatively low cost. As a consequence, cross-sectional data can show an insignificant or even negative relation between crisis length and resolution cost, even if such a true positive relationship exists, provided resolution costs do not increase very rapidly with delays.

This bias is illustrated in Figure 4. The lengths OA an OC measure different levels of initial crisis costs based on banking system losses at the perceived start of a banking crisis. Crises with higher initial costs tend to deteriorate faster than low initial cost crises because more banks are likely to be in a lower net worth position, creating stronger incentives within the banking system for gambling behavior.11 The marginal cost of delay in high initial cost crises will be relatively high. In Figure 4 the slope of CD is greater than the slope of AB.

Figure 4:
Figure 4:

Downward Bias in the Observed Relationship Between Crisis Cost and Length

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

Bank regulatory authorities, therefore, will have a stronger incentive to cut short crises with higher initial costs. Observations of high initial cost crises will be clustered in, say, the oval on line CD, while low initial cost crises are observed to be bunched in the oval on line AB. The estimated relationship from all the observations, however, is shown by the dashed line EF, which produces a downwardly biased estimate of the correct underlying marginal cost of delay in dealing with banking crises.

Given the moral hazard problem, authorities should try to minimize the action lag and take resolution measures that are commensurate with the size of the problem at the end of the perception lag. The marginal cost attributable to the delay in taking appropriate action should be the difference between the costs at the end of the perception lag and the total costs of resolution. This cost is the combination of the cost of delaying any action after getting the first impression of a crisis and the costs of taking insufficiently strong actions that do not succeed in stabilizing the situation, even if these actions are taken in a timely manner.

Typically, however, information is not available that allows the resolution costs of a banking crisis to be accounted accurately between those that were realized before the end of the perception lag and those taken on later. Even going back to the contemporaneous record at the end of the perception lag, as it may appear in newspaper accounts or statements of public officials, may misstate the true cost at that time, since contemporaneous estimates are unlikely to be based on full information.

Incorporating determinants of initial costs

Given this information problem, an alternative approach is to include determinants of the size of the banking problem at the end of the perception lag in the estimation of the relation between cost and length of the banking crisis. These determinants would account for cross-country variation in the initial size of the banking crisis and allow a more accurate estimate of the marginal effect of delay on resolution cost.

Recent studies point to a broad variety of potential determinants of banking crises.12 This paper selects only a limited set of likely macroeconomic factors—credit growth, economic activity, and real interest rates—in order to preserve degrees of freedom in the small sample and to limit the number of countries excluded because of missing data.13 The literature generally contends that banking crises are preceded by credit booms, economic weakness, and a rise in real interest rates that may be a consequence of financial liberalization.14 These three factors are commonly highlighted as macroeconomic determinants of banking distress but they are hardly exhaustive. The purpose of this paper, however, is not to test comprehensively for determinants of banking crises. Rather, the purpose is to introduce some generally accepted variables that may account for cross-country variation in the perceived costs of the crisis at the end of the perception lag in order to estimate better the effect of delay on resolution costs. The credit boom measure is the average of the difference between the ratio of domestic credit to GDP and a 10-year linear trend value of that ratio for the two years preceding the onset of the banking crisis. The activity variable is the average ratio of the output gap (measured against a 10-year linear trend for real GDP) for the two years preceding the banking crisis.

The real interest rate measures the average difference between real interest rates and the 10-year linear trend of real interest rates for the two years preceding the banking crisis. Real interest rates were calculated using bank lending rates, if available, or other bank rates or money market rates, if not. Inflation rates were calculated from implicit GDP deflators. Resolution costs of banking crises are assumed to increase directly with the credit boom and interest rate measures and inversely with the output measure.

The result of incorporating these macroeconomic determinants into the cross-country estimation of resolution costs is presented in equation (3).

Eq.(3)cost=7.60(1.38)+0.84(0.81) length+0.92(1.94)*credit+2.87(2.46)*output+0.02(1.41) Real interest rate

*significant at 5 percent confidence level

R2 =.32 Degrees of freedom = 20

T-stats in parentheses

The inclusion of the macroeconomic determinants in the estimation improves the general fit of the equation; the R2 statistic increases to .32. The coefficients on credit acceleration and the output gap are significant, but the output gap has the wrong sign. The coefficient on the length term increases (to 0.84 from 0.24 in equation (1)) but remains insignificant. The wrong sign on the output gap and the omission of other potential determinants point to remaining problems of misspecification and mismeasurement. These problems may account for the continuing insignificance of the length term. Still, for the CK sample, which offers the largest available data set on both the cost and length of banking crises globally, the evidence does not point to delay in resolution actions having a clear-cut strong effect on resolution costs.

IV. Crisis Length and Growth

As noted earlier, resolution costs measure the fiscal costs of a banking crisis but are subject to various errors and do not incorporate various indirect costs to the government or general costs to the economy. An alternative approach, then, to examining whether longer crises are more costly is to look at their effects on “bottom-line” indicators of economic performance.

Economic growth is a natural final performance indicator. Banking crises can depress economic growth in various ways. Concern about the condition of the banking system can restrict sources of funds and raise their cost. In reaction banks not facing an incentive to gamble may limit the amount of credit available and raise borrowing costs to the general economy, slowing growth. In addition, uncertainty about the costs to the government of bank losses and any ensuing tax liabilities may depress aggregate demand.

If longer crises have greater economic costs, they should act as a greater burden on economic growth. A simple test of this thesis is to see whether any shortfall between economic growth in the crisis period and some trend measure of growth is positively correlated with crisis length.

Such a measure was constructed, where data were available, by subtracting average growth during the crisis period from average growth during the 10 years preceding the crisis.15 The observations are presented in Table 6 and Figure 5. They indicate that short-lived crises of one-year duration are not associated with a growth shortfall. Episodes of 2 to 7 years are correlated with substantial growth shortfalls, although not in a monotonic fashion. Beyond 7 years in length the association breaks down and the number of observations is scarce.

Figure 5:
Figure 5:

Average Growth Shortfall vs. Length of Banking Crisis

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

Table 6.

Growth Shortfall by Crisis Length (CK Sample)

article image

See Appendix 3 for full sample data.

Simple regressions of the growth shortfall on crisis length show only a marginally stronger relationship than was the case using resolution costs. Figure 6 shows the fitted trend line indicated in Equation (4). Again there is a positive relation between crisis length and cost—this time measured by the growth shortfall—but it is statistically weak.

Figure 6:
Figure 6:

Growth Shortfall vs Length of Banking Crisis

Citation: IMF Working Papers 1999, 030; 10.5089/9781451844894.001.A001

* significant at 5 percent confidence level

Long-lived crises are very uncommon — only one or two are observed at each length above 5 years. Consequently, confidence in the information value of these scanty observations will be relatively low. Excluding the unusual events of very long-lived crises (more than 5 years) on such concerns about measurement error in the sample observations does improve the results, as shown in Equation (5); crisis duration then has a significant, as well as a more sizeable, effect on bank costs. These results suggest that part of the difficulty in obtaining a significant relationship between crisis length and resolution cost is due to the variety of measurement problems that plague resolution costs.

Using a growth shortfall as a measure of overall cost raises the possibility of a causal linkage in the opposite direction: a growth shortfall may produce a banking sector crisis as the consequence of a generally weaker economy. Indeed, that association is likely to be stronger, the longer the duration of the shortfall, since prolonged economic weakness is needed to destroy banking system capital and create a crisis. If causation runs from economic weakness to bank crises, though, the incidence of crises by duration is more likely to be positive, since periods of prolonged growth shortfall are more likely to produce banking crises. In that case, more crises would be observed at longer crisis lengths than at shorter lengths. The incidence of crises by duration in the sample, however, is negative, which is consistent with causality running from crisis length to growth shortfall.

V. Conclusions

Recent studies of banking crises show important differences regarding the timing, duration and cost of the crises. An analytical consensus on the features is still elusive. Dating banking crises is a similar problem to demarcating recessions; it involves a necessary degree of expert judgment. As is the case with recessions, an analytical consensus can be achieved by relying on the judgment of a group of credible experts as arbiters.

Data from the recent studies also allow analysis of the relation between crisis length and measures of crisis cost. Arguments based on the incentives created from the moral hazard of delay in resolving a banking crisis suggest that a positive relation should exist between crisis length and cost. Tests using a cross-section of available data do not show a significant positive relation. Regulators have a clear incentive to more quickly resolve crises where delay threatens to be more costly. Crises where bank losses from initial shocks are high create the risk of further relatively high losses from delay. But even after trying to account for this effect by including determinants of initial crisis costs, a significant relation between crisis length and cost remained elusive. Results are somewhat better if costs are measured by forgone GDP growth rather than by budgetary resolution costs, possibly because resolution costs are a relatively narrow cost concept prone to a variety of measurement errors.


Episodes of Banking Crises

article image
article image
article image

Highlighted areas represent episodes of bank crises that have definite periods associated with them.

Percentage of GDP, unless otherwise indicated.

Sources: Caprio and Klingebiel (1996b), Demirguc-Kunt and Detragiache (1998), Dziobek and Pazarbasioglu (1995), Kaminsky and Reinhart (1996), Lindgren, Garcia and Saal (1996).