11.07.2015 Views

“Rules of Thumb” for Sovereign Debt Crises

“Rules of Thumb” for Sovereign Debt Crises

“Rules of Thumb” for Sovereign Debt Crises

SHOW MORE
SHOW LESS
  • No tags were found...

Transform your PDFs into Flipbooks and boost your revenue!

Leverage SEO-optimized Flipbooks, powerful backlinks, and multimedia content to professionally showcase your products and significantly increase your reach.

- 3 -<strong>of</strong> short-term debt may need IMF support to avoid a liquidity run or “roll-<strong>of</strong>f” crisis. Conversely, highlyindebted countries may face a debt crisis, unless there is a strong and credible fiscal consolidation. Inaddition, it is argued that conditionality should set targets indicating that a country’s macroeconomicfundamentals are heading towards a relatively “safe” zone. In the paper these concepts <strong>of</strong> liquidity crisis,insolvency crisis, crisis triggered by weak macro-fundamentals, and relatively “safe zone” are madeprecise. Unless the diagnosis is correct, it is hard to get the policy cure right.This empirical analysis is based on a dataset containing annual observations <strong>for</strong> 47 emerging marketeconomies from 1970 to 2002. A country is defined to be in a state <strong>of</strong> “debt crisis” if it is classified asbeing in default by Standard & Poor’s, or if it receives a large non-concessional IMF loan (where “large”means in excess <strong>of</strong> 100 percent <strong>of</strong> IMF quota). Standard & Poor’s rates sovereign issuers in default when agovernment fails to meet principal or interest payment on an external obligation on due date (includingexchange <strong>of</strong>fers, debt equity swaps, and buy back <strong>for</strong> cash).We employ the Classification and Regression Tree methodology (CART) <strong>for</strong> classification andprediction. 3 CART is a computer-intensive data mining technique that selects explanatory variables, theircritical values, and their interactions in order to identify “safe” from “crisis-prone” types. The mainconclusions <strong>of</strong> our empirical analysis are as follows.First, out <strong>of</strong> 50 candidate variables, 10 predictor variables turn out to be sufficient <strong>for</strong> classification andprediction: total external debt/GDP ratio; short-term debt reserves ratio; real GDP growth; public externaldebt/fiscal revenue ratio; CPI inflation; number <strong>of</strong> years to the next presidential election; U.S. treasurybills rate; external financial requirements (current account balance plus short-term debt as a ratio <strong>of</strong><strong>for</strong>eign reserves); exchange rate overvaluation; and exchange rate volatility.Second, a relatively “safe” country type is described by a handful <strong>of</strong> economic prerequisites: low totalexternal debt (below 49.7 percent <strong>of</strong> GDP); low short-term debt (below 130 percent <strong>of</strong> reserves); low3 The analysis employs the data mining s<strong>of</strong>tware CART developed by Sal<strong>for</strong>d Systems.


- 4 -public external debt (below 214 percent <strong>of</strong> fiscal revenue); and an exchange rate that is not excessivelyover-appreciated (overvaluation below 48 percent).Third, three major types <strong>of</strong> risks are identified: (i) solvency (or debt unsustainability); (ii) illiquidity; and(iii) macro-exchange rate risks. The debt unsustainability risk types are characterized by: external debt inexcess <strong>of</strong> 49.7 percent <strong>of</strong> GDP, together with monetary or fiscal imbalances, as well as by large externalfinancing needs that signal illiquidity as an element <strong>of</strong> debt unsustainability. Liquidity risk types areidentified by moderate debt levels, but have short-term debt in excess <strong>of</strong> 130 percent <strong>of</strong> reserves coupledwith political uncertainty and tight international capital markets. Macro-exchange rate risk types arisefrom the combination <strong>of</strong> low growth and relatively fixed exchange rates. Each <strong>of</strong> these risk types differ intheir likelihood <strong>of</strong> producing a crisis. Finally, our model has excellent predictive capacity in sample, whilethe out-<strong>of</strong>-sample <strong>for</strong>ecast have both less false alarms and less correct predictions than the Early WarningSignal (EWS) literature.The analysis has one important, albeit simple, implication <strong>for</strong> sustainability analysis. It shows thatunconditional thresholds, <strong>for</strong> example <strong>for</strong> debt-output ratios, are <strong>of</strong> little value per se <strong>for</strong> assessing theprobability <strong>of</strong> default. One country may be heavily indebted but have a negligible probability <strong>of</strong> default,while a second may have only moderate values <strong>of</strong> debt ratios while running a considerable default risk.Why? Because the joint effects <strong>of</strong> short maturity, political uncertainty, and relatively fixed exchange ratesmake a liquidity crisis in the latter much more likely than a solvency crisis in the <strong>for</strong>mer, particularly if thelarge external debt burden goes together with monetary stability, a large current account surplus, andsound public finances.The plan <strong>of</strong> the paper is the following. Section II contains a review <strong>of</strong> the literature. Section III describesthe dataset. The Classification Tree methodology is reviewed in Section IV, and applied to the data inSection V. Section VI discusses an interpretation <strong>of</strong> the results. Section VII and VIII contain standardregression analysis and a robustness check based on a generalization <strong>of</strong> the CART procedure. Thepredictive power <strong>of</strong> the model is discussed in Section IX, and Section X per<strong>for</strong>ms some “rolling tree”exercises in order to understand if the crises <strong>of</strong> the nineties are “structurally” different from those in the


- 5 -previous decades. The final Section XI discusses some extensions and summarizes the main conclusionsand policy implications.II. Review <strong>of</strong> the LiteratureMost empirical studies consider that countries may either be unable to repay their debt, because they areinsolvent or illiquid, or that they may be unwilling to repay their debt, based on a consideration <strong>of</strong> therelative costs and benefits <strong>of</strong> default. Thus, these studies suggest proxies <strong>for</strong> ability and willingness torepay as the determinants <strong>of</strong> debt crises.Different studies use different crisis definitions, depending on the specific research question and thein<strong>for</strong>mation available in the data source used. A priori, there is no single empirical definition <strong>of</strong> whatshould constitute a sovereign default or a debt crisis. Some studies compile a list <strong>of</strong> debt crisis or defaultepisodes from case studies and anecdotal evidence (e.g., Beers and Bhatia, 1999; or Beim andCalomiris, 2001). Others rely on a quantitative approach. For example, Detragiache and Spilimbergo(2001) define a country to be in a debt crisis if the country has arrears on external obligations towardscommercial creditors in excess <strong>of</strong> 5 percent <strong>of</strong> commercial debt outstanding, or has a rescheduling orrestructuring agreement with commercial creditors. This definition does not differentiate betweensovereign or private sector arrears and/or rescheduling due to data limitations. Another problem <strong>of</strong> thisquantitative definition is that it might exclude some incipient debt crises that were only avoided by largescalefinancial support from <strong>of</strong>ficial creditors (International Financial Institutions (IFIs)), and/or frombilateral creditors. A data source that provides a uni<strong>for</strong>mly compiled in<strong>for</strong>mation on sovereign default isStandard & Poor’s (2002) that defines a country to be in default as long as the sovereign is not current onany <strong>of</strong> its debt obligation.Empirical studies <strong>of</strong> the determinants <strong>of</strong> debt crisis are closest in nature to an Early Warning Signal model.Factors influencing the probability <strong>of</strong> a debt crisis are identified by means <strong>of</strong> probit /logit regressions or asignal model. Most studies have focused on the debt crisis <strong>of</strong> the 1980s, but some recent ef<strong>for</strong>ts have


- 6 -looked at crises occurring in the 1990s. 4 Reinhart (2002) finds that in 84 percent <strong>of</strong> the cases in hersample, a debt crisis is preceded by a currency crisis. Detragiache and Spilimbergo (2001) find that shorttermdebt is endogenous to the model, as countries find it more and more difficult to borrow long term inthe run-up to a debt crisis. While most studies use macroeconomic variables only in levels, Catão andSutton (2002) also include measures <strong>of</strong> volatility in trade, fiscal policy, monetary policy and exchangerate policy. Manasse, Roubini, and Schimmelpfenning (2003) estimate a logit model <strong>of</strong> sovereign debtcrisis that includes a large set <strong>of</strong> emerging market economies <strong>for</strong> the 1970–2002 period; thus, they includethe sovereign crises <strong>of</strong> the last decade in their sample. Their model predicts about three quarters <strong>of</strong> allcrises entries while sending few false alarms.Taken together, the existing literature suggests several factors that are at the core <strong>of</strong> an empirical modelattempting to predict sovereign default: measures <strong>of</strong> solvency, such as public and external debt relative tothe capacity to pay; liquidity measures such as short-term external debt and external debt service, possiblyas a ratio <strong>of</strong> <strong>for</strong>eign reserves or exports; political, institutional and other variables capturing a country’swillingness to pay; macroeconomic variables such as real growth, inflation, exchange rate, etc., capturingboth ability to pay and willingness to pay; measures <strong>of</strong> external volatility and volatility in economicpolicies. Thus, we use these variables and measures in the empirical study.While the tree methodology that we employ has been used in a limited number <strong>of</strong> economic studies andeven applied to the case <strong>of</strong> currency crises (see Ghosh and Ghosh, 2002, and Frankel and Wei, 2004 ) noprevious study has used this methodology to assess the determinants <strong>of</strong> sovereign crises and to predictthem.III. Data and DefinitionsThe full dataset includes annual in<strong>for</strong>mation on 47 economies with market access from 1970 to 2002(Table 1). 5 A country is defined to be in a debt crisis if it is classified as being in default by Standard &4 See <strong>for</strong> example Detragiache and Spilimbergo (2001) <strong>for</strong> a study including recent episodes.5 The debt crisis indicator is derived from data provided by Standard & Poor’s and data on IMF lending. Data onexternal debt and public debt is taken from the World Bank’s Global Development Finance database (GDF), as wellas from IMF sources. Data on public finance and other macroeconomic variables are taken from the IMF’s WorldEconomic Outlook database, as well as the Government Finance Statistics database (GFS).For transition economies,the sample period is 1995–2002. Not every variable is available <strong>for</strong> all countries or <strong>for</strong> the full-time period.


- 12 -weight differently the cost <strong>of</strong> misclassifying crisis and non- crisis episodes 11 : a higher cost <strong>of</strong> missing acrises raises p 1 (t) and similarly increases the frequency a crisis is called. The option <strong>of</strong> choosing priorsand misclassification costs at the outset (i.e., be<strong>for</strong>e running the CART algorithm) influences the modelchoice (i.e., the set <strong>of</strong> indicators and thresholds). This is a key difference between CART and standardEWS. For example, Berg, Borensztein, and Pattillo (2004) use a misclassification cost function (weightingType I and Type II errors) to identify the probability cut-<strong>of</strong>f point that would best predict crisis and noncrisisevents out-<strong>of</strong>-sample only after having estimated the probit model. As a result, their cost-functioncannot influence the choice <strong>of</strong> the indicators and coefficients in the probit model 12 . Here, the choice <strong>of</strong>priors (and costs) shapes the model, as well as determines the cut-<strong>of</strong>f point <strong>for</strong> labeling terminal nodes ascrisis or non-crisis types (see below).In order to decide when to stop the growing process, the algorithm calculates the gain from furthersplitting in terms <strong>of</strong> the reduction in the rate <strong>of</strong> misclassification from each parent to the two child nodes.A standard approach would be to stop growing the tree when the reduction in the misclassification ratebecomes lower than the change in the penalty associated with the number <strong>of</strong> terminal nodes. However,such an approach would not be capable <strong>of</strong> discovering possibly interesting splits further down the tree.Thus, CART initially grows a maximal, over-fitted, tree, and then proceeds to “pruning” branchesbackward. The best tree ( in terms <strong>of</strong> “goodness <strong>of</strong> fit”) is determined by looking <strong>for</strong> the tree whosemisclassification error rate (adjusted <strong>for</strong> a penalty on the number <strong>of</strong> nodes) is lowest.Finally, the criterion to label a terminal node as crisis or non-crisis-type is quite intuitive: classify terminalnode t as crisis-type if p 1 (t) > p 0 (t). Note, from equation (2), that when the prior distribution coincideswith the empirical likelihood, π 1 = N 1 /N , then p 1 (t)=N 1 (t)/N(t) and the criterion yields the plurality rule:11 One can explicitly set different costs from misclassifying a crisis (C1 ) and a non crisis (C 0) observation. Doing so modifies the<strong>for</strong>mula <strong>for</strong> the (adjusted) posterior probabilities as follows:C1π1N1( t ) / N1∑ Cjπjj = 0 , 1p1( t ) =Ciπi∑Ni( t ) / NiC πi = 0, 1∑j = 0 , 1jjThe relative misclassification costs and prior probabilities can be used inter-changeably in order to make the crisis classificationmore conservative, given the subjective preference <strong>for</strong> reducing the chance <strong>of</strong> missing crises.


- 13 -classify node t as type 1 if the node contains more type-1 observations than type 0, N 1 (t)>N 0 (t).Conversely, if priors are “agnostic” in the sense that π 1 = π 0 = ½, then it is easy to see that node t will belabeled as type-1 provided the these types are relatively more abundant in the node than in the entiresample: N 1 (t) / N 0 (t) > N 1 / N 0 . From the <strong>for</strong>mulas above it descends that the posterior probability <strong>of</strong> acrisis, p 1 (t) , is increasing in the relative cost <strong>of</strong> a crisis C 1 and in the prior probability π 1. Thus the largerthese two parameters, the more likely it becomes that CART will label a node as crisis prone, and thelarger the resulting (Type II) error <strong>of</strong> false alarms.For the classification tree presented below, we have used the following specifications: (i) the criterion tobe minimized is the Gini index in equation (1); (ii) the a priori distribution <strong>of</strong> crises is taken to be equal tothe sample distribution, π i = N i / N; (iii) the cost <strong>of</strong> missing a crisis is set to 7 times the cost <strong>of</strong> missing anon-crisis (which is normalized to 1). From the expression in the previous footnote, these parameters yielda posterior probability <strong>of</strong> a crisis in node t, p 1 (t )= 7N 1 (t)/[N(t) + 6N 1 (t)]. Thus, a node is labeled as crisis(p 1 (t)>1/2 ) when it contains at least 12,5 percent <strong>of</strong> crises (that is when N 1 (t)/N(t) > 0,125). This cut-<strong>of</strong>fpoint takes into account the fact that in our sample crises episodes are possibly over-represented withrespect to the “true” population (they make up 20,5 percent <strong>of</strong> sample observations), so that we want to“call” a crisis relatively more frequently. Our findings, however, are robust to this choice <strong>for</strong> the cut-<strong>of</strong>fpoint <strong>for</strong> calling a node a “crisis” type (see below).CART FeaturesThe CART algorithm displays a number <strong>of</strong> interesting features. First, it is well suited <strong>for</strong> discoveringcontext dependence, interactions and heterogeneity. Unlike parametric models, which aim at uncovering asingle dominant (typically linear) structure in the data, CART is designed to work with data that mighthave multiple structures. In fact, one <strong>of</strong> our most interesting findings is exactly that not all crises are thesame. The methodology is robust to the presence <strong>of</strong> outliers among the independent variables. These donot generally affect CART because splits usually occur at non-outlier values. This feature is particularlyuseful when dealing with emerging markets, where, particularly in crisis times, variables such as inflation12 See Chamon, Manasse and Prati (2006)


- 14 -and exchange rate depreciation take extraordinary values. No specification search is necessary. Thisfeature is useful when theory does not precisely identify the “right” explanatory variable (which indicator<strong>of</strong> liquidity constraints should we use: short-term debt over GDP, over exports, reserves, or over totaldebt?). Moreover, the algorithm yield the same results <strong>for</strong> any monotone trans<strong>for</strong>mation <strong>of</strong> the variables(<strong>for</strong> example, levels or logarithms). The procedure is non-parametric, as it does not require specification <strong>of</strong>a functional <strong>for</strong>m <strong>for</strong> the exogenous variables. Also, missing values <strong>for</strong> predictors are handled veryeffectively, 13 so that CART can be used also with data sets that have a fraction <strong>of</strong> missing values, as <strong>of</strong>tenis the case, <strong>for</strong> example, with fiscal data <strong>for</strong> developing countries. Finally, the algorithm is designed toavoid “over fitting” the model to the data, so that predictions should remain accurate when applied to freshdata. When data limitations do not allow <strong>for</strong> having a separate test sample, CART proceeds by dividing thesample into 10 roughly equal parts, each containing a similar distribution <strong>for</strong> the dependent variable. Ittakes the first nine parts <strong>of</strong> the data, constructs the largest possible tree, and uses the remaining 1/10 <strong>of</strong> thedata to obtain initial estimates <strong>of</strong> the misclassification error rate <strong>of</strong> the sub-trees. The same process is thenrepeated (growing the largest possible tree) on another 9/10 <strong>of</strong> the data while using a different 1/10 part asthe test sample. The process continues until each part <strong>of</strong> the data has been held in reserve one time as a testsample. The results <strong>of</strong> the ten mini-test samples are then combined in evaluating the optimal (lowestadjusted misclassification) tree. This procedure, called cross-validation, implies that the algorithm strivesat finding a model that per<strong>for</strong>ms well on completely fresh data, even in the absence <strong>of</strong> an independent testsample.CART has a few shortcomings, however. Unlike regression or probit/logit analysis, the individualmarginal contribution <strong>of</strong> each variable to the probability <strong>of</strong> belonging to a class cannot be ascertained:CART assigns a single probability to all cases belonging to the same node. Furthermore, the procedure isnot well suited <strong>for</strong> uncovering “general” relationships that hold across the whole sample: a linear effect13 For each split in the tree, CART develops alternative splits (surrogates), that is, variables that produce a similarallocation <strong>of</strong> observations in child nodes. These variables are used when the primary splitting variable is missing.Thus, a missing value does not imply that all the observations need to be thrown away, as in regression analysis. Themissing value is replaced by the best surrogate.


- 15 -would show up as the same variable being selected repeatedly with different thresholds. Moreover, as thepattern <strong>of</strong> discovery becomes progressively more localized, sample-wide in<strong>for</strong>mation is not used. 14 Third,since CART is non-parametric and free from assumptions on probability distributions, confidence intervalscannot be meaningfully attached to the threshold values. However, one can estimate a logit model withdummy variables corresponding to node inequalities, and test <strong>for</strong> their significance (see below). Finally, ifa variable is even slightly outper<strong>for</strong>med by another as a split, it may never appear in the final tree, even ifit is more closely associated with class membership than other variables appearing down in the tree. Thus,one may incorrectly deduct that the omitted variable is not “important.” This problem (called “masking”)is somewhat akin to having two significant regressors that happen to be strongly collinear: one drops outfrom the regression. The consequence <strong>for</strong> classification is that the variables appearing in a tree may besensitive to changes in the sample, in the list <strong>of</strong> variables considered, and/or in the a priori distribution. Apossible way out is to rank each variable by looking at the potential effect <strong>of</strong> the variable on classification.This procedure explicitly accounts <strong>for</strong> a variable’s ability to classify observations at each node, even when“masked” by the first-choice split. It yields a measure (the Gini index <strong>of</strong> “importance”) that does notdepend on the variables actually appearing in the tree (more on this below).The problem <strong>of</strong> robustness to the addition <strong>of</strong> new variables and observations is rather complex, however.If a new variable is in<strong>for</strong>mative enough to replace a pre-existing variable, particularly in the top splits,there is a good chance that the branches developing from there onwards will feature a somewhat differentset <strong>of</strong> variables and thresholds. Similarly, the introduction <strong>of</strong> additional years or countries to the samplemay lead to substantial changes in the optimal tree. We will discuss and apply a solution based on MonteCarlo simulations proposed by Breiman (1992) in Section VIII below, where we address the issue <strong>of</strong> therobustness <strong>of</strong> our results.VI . The Empirical Tree14 Related to this is the criticism that the procedure is only “one-step” optimal and not “overall” optimal. For example,suppose that the procedure produces a tree with ten terminal nodes. If one could search all possible partitions <strong>of</strong> thedataset in ten terminal nodes <strong>for</strong> the one partition that minimizes the Loss function , the two result may be quitedifferent (Breiman and others, 1984).


- 16 -CART selects the following 10 variables out <strong>of</strong> the 50 candidates listed in Table 2: total external debtin percent <strong>of</strong> GDP; short-term debt relative to <strong>for</strong>eign reserves; public external debt in percent <strong>of</strong> publicrevenues; real GDP growth; inflation; the U.S. treasury bill rate; exchange rate overvaluation; exchangerate volatility; the ratio <strong>of</strong> external financing requirements to <strong>for</strong>eign reserves; and the number <strong>of</strong> yearsbe<strong>for</strong>e a presidential election.The first rule splits the sample into two branches (see Figure 1) [insert Figure 1 here] : (i) episodes withhigh external debt (more than 49.7 percent <strong>of</strong> GDP) go to the right—here the conditional crisis likelihoodrises from 20.5 percent in the entire sample to 45.4 percent; and (ii) episodes with low external debt to theleft—with default likelihood <strong>of</strong> 9.7 percent. Episodes <strong>of</strong> high debt (more than 49.7 percent <strong>of</strong> GDP) arefurther split into high/low inflation (above/below 10.5 percent). The <strong>for</strong>mer incur the largest defaultfrequency, 66.8 percent, see terminal node 14. More than half <strong>of</strong> all the crisis episodes in the samplesatisfy these two simple conditions. For example, the high external debt plus high inflation criterion wasmet one year ahead <strong>of</strong> the crises in Jamaica, Egypt, Bolivia, Peru, Ecuador, Uruguay, Indonesia, Bolivia,Morocco, Turkey, South Africa, Uruguay, Brazil, and Venezuela. Terminal node 7 is second in terms <strong>of</strong>number <strong>of</strong> crisis episodes. How do we get there? Despite intermediate external debt levels (between49.7 percent and 19 percent <strong>of</strong> GDP), the joint effect <strong>of</strong> short-term debt (exceeding 130 percent <strong>of</strong>reserves), political uncertainty (less than five years to the next presidential elections), and relatively rigidexchange rates (low volatility) conjure to raise the crisis likelihood to 41 percent.By contrast, going down from the top to the left, towards terminal node 3, one finds the circumstancesthat are more favorable <strong>for</strong> reducing risks: low external debt, low short-term debt to reserves on aremaining maturity basis (below 130 percent) and low public external debt to revenue (below210 percent), coupled with the economy not being in recession. Under these circumstances, the likelihood<strong>of</strong> being in a crisis is just 2.3 percent. About 58.4 percent <strong>of</strong> all non-crisis episodes satisfy theseconditions.Based on the set <strong>of</strong> rules <strong>of</strong> this tree, the observations can be classified as crisis-prone or not crisis-prone.As explained previously, a particular final node is classified as crisis-prone (not crisis-prone), if the within


- 17 -node posterior probability <strong>of</strong> crisis is higher (lower) than ½ . As discussed in the methodology section, ourchoice <strong>of</strong> parameters implies a cut-<strong>of</strong>f share <strong>of</strong> within-node crisis frequency <strong>of</strong> 12,5 percent, that is, belowthe proportion <strong>of</strong> crises in the sample, 20.5 percent. As can be seen from the figure, the largest percentage<strong>of</strong> (missed) crises in “non-crisis” terminal node is only 2.3 percent, while the smallest percentage <strong>of</strong> crisesin a “crisis” node is 40 percent. Thus, the tree separates the two types quite effectively, so that any cut-<strong>of</strong>fthreshold <strong>for</strong> labeling nodes, set between 2.3 percent and 40 percent, would not affect the finalclassification.The tree has one particularly important implication <strong>for</strong> sustainability analysis. It shows that unconditionalthresholds <strong>for</strong> debt output ratios are <strong>of</strong> little value per se <strong>for</strong> assessing the probability <strong>of</strong> default. Taketerminal nodes 7 and 11. The <strong>for</strong>mer has only moderate values <strong>of</strong> debt ratio, between 49.7percent and 19 percent, but the likelihood <strong>of</strong> a crisis is high, 41 percent. The latter has a debt ratio <strong>of</strong> atleast 49.7 percent, but the crisis probability is just 2 percent. Why? Clearly, other factors are at play. Whatmakes node 7 risky is the compound effect <strong>of</strong> short maturity <strong>of</strong> the debt, political uncertainty andrelatively fixed exchange rates. What makes node 11 safe, despite the large debt burden, is monetarystability, a large current account surplus, and relatively large fiscal revenues that guaranteed solvency onpublic debt.Classification TableTable 3 shows some interesting in<strong>for</strong>mation on the classification tree. [insert table 3 here]The first column lists the terminal node’s number. The second column reports the set <strong>of</strong> inequalitiessatisfied by each node’s observations: <strong>for</strong> example, row 14 identifies terminal node 14, containing all thecases (country-year) where total external debt/GDP exceeds the threshold <strong>of</strong> 49.7 percent and inflationexceeds 10.5 percent. Column 3 counts the number <strong>of</strong> observations that satisfy the node criteria, and tellsus how “relevant” is a terminal node. For example, the table shows that the three most populated nodes arenumber 3 (607 observations), 14 (196 cases) and number 7 (130 episodes). Thus, <strong>for</strong> example there are196 observations (in node 14 ) that satisfy the two inequalities on debt and inflation. The fourth columnreports the empirical likelihood <strong>of</strong> a crisis, conditional on the node’s inequalities: <strong>for</strong> example, the


- 18 -likelihood <strong>of</strong> a crisis next year <strong>for</strong> a country with high debt and inflation (node 14), e.g. the proportion <strong>of</strong>observations satisfying the two inequalities, which had a crisis the following year is 66.9 percent. Thisnode is classified as a crisis node (see column 8) since the crisis (posterior) probability exceeds ½. 15 Thisallows one to calculate the Type II error. In the case <strong>of</strong> node 14, 33.1 percent <strong>of</strong> the cases displaying highdebt and inflation turned out not to experience a crisis the following year, and were there<strong>for</strong>e misclassifiedas crises. The fifth and sixth columns give in<strong>for</strong>mation on how frequent the identified characteristics arefound among crises and non-crises. For example, looking at node 14 and column 5, we see that more thanhalf (50.2 percent) <strong>of</strong> all sovereign debt crises in the sample satisfy the debt and inflation criteria. Thisnumber is computed as the ratio <strong>of</strong> the crises in the node to the total number <strong>of</strong> crises in the sample.Similarly, the sixth column shows how frequent are the reported characteristics among non-crisisobservations: <strong>for</strong> example, from the row 14, column 6, we see that only 6.4 percent <strong>of</strong> all non-crisisepisodes displayed large external debt and inflation. Finally, we want to derive a measure <strong>of</strong> “how safe”are the thresholds <strong>for</strong> fundamentals. To do so, we divide the percentage <strong>of</strong> non crises in a node, N 0 (t)/N 0 ,by the share <strong>of</strong> observations in the node, N(t)/N (<strong>for</strong> node 14 we divide column 6 by (196/1276)). Theresult is shown in the seventh column <strong>of</strong> the table. This “safety” index that takes value greater than onewhen the node is “safer” than the overall sample, and smaller when it is “riskier”.InterpretationThese calculations lend themselves to an interesting interpretation. We can regroup the terminal nodes int<strong>of</strong>our blocks, which identify different types <strong>of</strong> macroeconomic scenarios. Block A can be interpreted asdescribing the characteristics <strong>of</strong> the “(relatively) safe fundamentals” type: observations in this block areclassified as non-crisis prone. Blocks B, C, and D identify potential defaulters, which are subject to risks<strong>of</strong> different sorts. Observations therein are classified as crisis prone.15 As explained above the posterior probability does not in general coincide with the empirical likelihood reported inthe table, since it reflect the choice <strong>of</strong> priors and cost weights. For example, applying the <strong>for</strong>mula p 1 (t) = 7N 1 (t)/[N(t)+ 6 N 1 (t)] <strong>of</strong> footnote 11 gives a posterior probability <strong>of</strong> crisis in node 14 <strong>of</strong> 93,4 per cent which exceed the likelihood(66,9 percent) due to the high relative cost <strong>of</strong> missing a crisis (7:1).


- 19 -“Relatively Safe” Fundamentals, type A. Nodes 1, 4, 6, 8, 9, 11, and 3, in descending order <strong>of</strong> “safety,”show the prerequisites <strong>for</strong> a relatively risk-free environment. Low total external debt (below 49.7 percent)is the common denominator <strong>of</strong> these nodes (with the exception <strong>of</strong> partition 11). Low short-term debt overreserves (below 130 percent), low public external debt over revenue (below 215 percent), low inflation(below 10.5 percent) and not too strong recession (growth above -5.5 percent) also characterize many (butnot all) <strong>of</strong> these partitions. By far the more representative partition is node 3 that contains 607 episodesamounting to 58 percent <strong>of</strong> all tranquil nodes. Node 11 comes next in terms <strong>of</strong> observations. Node 11shows that low external debt is by no means necessary <strong>for</strong> avoiding a crisis. Despite a large <strong>for</strong>eign debt, acountry may still enjoy a relatively safe environment (2 percent crisis likelihood), provided inflation isunder control (below 10.5 percent), external financial requirements over reserves are not too high (below140 percent) and public external debt over revenue is not too large (below 310 percent). Node 3 shows thatlow external debt is by no means sufficient <strong>for</strong> avoiding a crisis. Despite showing relatively soundfundamentals as defined in node 3, 5.4 percent <strong>of</strong> these observations nevertheless had a crisis thefollowing year (they are listed in column 9).The nature <strong>of</strong> the inequality constraints in the following “crisis prone” nodes suggests an intuitiveclassification into solvency (B), liquidity (C), macro-exchange rate (D) risks.Liquidity Crisis-Prone, type B is described by nodes 7 and 10. Despite relatively low or intermediateexternal debt ratios, these episodes share a ratio <strong>of</strong> short-term debt to reserves in excess <strong>of</strong> 130 percent.This feature, coupled with political uncertainty (presidential elections in less than five years) and relativelyfixed exchange rate (low volatility), in node 7, or with tight monetary conditions in international capitalmarkets (treasury-bill rate above 9.7 percent), in node 10, raise the likelihood <strong>of</strong> a crisis to about40 percent. Solvency risks <strong>of</strong> the first type (node 7) make up more than one-fifth (20.7 percent) <strong>of</strong> all debtcrisis episodes, while only 7.5 percent <strong>of</strong> observations with these characteristics turned out to be safe (the“false alarms”, see column 6)Unsustainable <strong>Debt</strong> Path (Solvency) Crisis-Prone, type C. Nodes 12, 13, 5 and 14 belong to this type. Inall these cases, either external debt exceeds 49.7 percent <strong>of</strong> GDP, or its public component exceeds


- 20 -215 percent <strong>of</strong> revenues. In partition 14, high total external debt, together with inflation above10.5 percent, raises the likelihood <strong>of</strong> a crisis from 20.5 percent (entire sample) to 67 percent. More thanhalf <strong>of</strong> all crises episodes in the sample satisfy these characteristics (see the fifth column). Only in 6.4percent <strong>of</strong> the cases where debt and inflation were above these thresholds a crisis did not occur in thefollowing year. Note that the high inflation rate could be here a proxy <strong>for</strong> fiscal problems: inability toreduce deficit may need to monetization <strong>of</strong> them and seignorage financing <strong>of</strong> them. In fact, terminal node5 is similar, but a solvency crises (4.2 percent <strong>of</strong> total) are signaled by public, rather than total, externaldebt and high inflation. In node 13 (12), high external debt and high financing external requirement (orfiscal imbalances measured by high public external debt over revenue), raise the default likelihood to47 percent (40 percent). These two nodes jointly account <strong>for</strong> another 15.7 percent <strong>of</strong> crises. The largeexternal financing need can be partly interpreted as an illiquidity problem: this financing need is highwhen the current account deficit is high and/or when there is a large amount <strong>of</strong> debt coming due; in eithercase, the country may be illiquid if creditors do not provide new financing and/or do not rollover theirdebt. Thus, crisis episodes in node 13 can be interpreted as cases where a country is not necessarily“insolvent” but rather has an unsustainable—and non-financeable—debt path given large stocks <strong>of</strong> debtand illiquidity measured by large financing needs.Exchange Rate and Macro-crisis prone, type (D) A small number <strong>of</strong> sovereign debt crises (1.5 percent) innode 2 appear to be driven by large exchange rate overvaluation (above 48 percent) and large recession(negative growth below -5 percent). These are the crises <strong>of</strong> El Salvador, 1981–82, Uruguay, 1983 (thecollapse <strong>of</strong> the Tablita, the <strong>for</strong>ward-looking crawling peg regime <strong>of</strong> some Southern Cone economies in theearly 1980s), and Venezuela, 1984.Finally, CART isolates a small number (13) <strong>of</strong> outliers in node 1: in these cases, the exchange rate was notover appreciated, values <strong>of</strong> external debt were small and there were no liquidity problems, and there wereno-crises despite a strong recession.How Different are <strong>Crises</strong> and how important Non-linearities?


- 21 -How different are the crisis types and how important are the threshold effects found in the analysis? Inorder to answer these questions, we can look at the within node summary statistics <strong>for</strong> the selected tencrucial variables <strong>of</strong> the tree. 16 It is useful to compare the most representative nodes <strong>of</strong> liquidity (7) and thesolvency crises (14 and 13) nodes. Indeed, the liquidity crisis node 7 has moderate average levels <strong>of</strong>external debt ( 37 percent <strong>of</strong> GDP), and <strong>of</strong> average public external debt (1.2 percent <strong>of</strong> GDP) , whileshort-term debt is very high, on average about 3 times larger than reserves. By contrast, solvency crisisnodes 13 and 14 not only display high short-term debt (2.7 times reserves), but also very high averagelevels <strong>of</strong> external debt (77 and 76 percent <strong>of</strong> GP respectively), and public external debt (2.4 and 3.2 timesrevenues). This suggests that solvency problems typically also entail liquidity difficulties, but notgenerally vice-versa. Solvency and liquidity crises have rather similar values <strong>for</strong> the other fundamentals.Finally, by comparing the characteristics <strong>of</strong> the “least safe” safe node 3, with the “safest” crisis pronenodes, 10 (and 12), it is possible to have an idea <strong>of</strong> the importance <strong>of</strong> threshold and non linear effects 17 .Nodes 3 and 10 differ mainly with respect to the ratios <strong>of</strong> short-term debt and external financialrequirements to reserves: these are substantially lower in the safe node,(0.6 against 2.1 <strong>for</strong> the averageshort-term debt over reserves, and 0.8 against 3.3 <strong>for</strong> external financing requirements over reserves). Thissuggests that non-linearities are not very sharp. Similarly, the safe node 3 and the risky node 12 differsubstantially in the average levels <strong>of</strong> external debt ( 27 and 110 percent <strong>of</strong> GDP) and public debt overrevenue (0.9 and 3.6 percent, respectively). Relative to node 10, the least safe “safe” node 3 is alsocharacterized by an exchange rate that is at the same time less overvalued and more flexible (i.e. morevolatile).16 For reasons <strong>of</strong> space, the data <strong>for</strong> all variables (total external debt, short-term debt over reserves, publicexternal debt over revenue, real growth, overvaluation <strong>of</strong> the exchange rate, inflation, years to the nextpresidential election, exchange rate volatility, US Treasury bill, external financing requirements) areavailable in a web appendix at the following address:http://www2.dse.unibo.it/manasse/Pdf/Web_Appendix_Within%20Node%20Summary%20statistics.pdf17 We thank one referee <strong>for</strong> this observation.


- 22 -In summary, our classification <strong>of</strong> crises captures quite well most <strong>of</strong> the episodes <strong>of</strong> the 1990s. Korea(1997), Mexico (1995), Brazil (1998) and (2001), are in the illiquidity node 7; and indeed these weretypical episodes <strong>of</strong> a liquidity crisis, not insolvency. Pakistan (1998) is also in this node: its debt levelswere not excessive but the country was illiquid given large lumpy debt servicing payments coming dueand its loss <strong>of</strong> market access. Pakistan was then <strong>for</strong>ced to restructure its external sovereign debt, and it didnot receive a large IMF package. Ukraine in 1998 (also in node 7) was a similar case <strong>of</strong> a solvent butilliquid country that coercively restructured its external debt by extending its maturities at below crisislevel market rates. Clear insolvency cases in node 14 are Ecuador (1999), which defaulted on its externaldebt, Indonesia (2002), with a very large debt burden (but it is not illiquid given its restructuring <strong>of</strong> manyexternal debt claims, both public and private), and Turkey (2000) with a very large debt burden. Turkeydid not default thanks to an exceptional IMF program, but based on this classification was borderlineinsolvent. 18 Russia (1998) and Argentina (1995) appear in the non-crisis node 3; hence the tree is unable topredict these crises. In the case <strong>of</strong> Argentina, the Tequila contagion from Mexico in 1994 played animportant role, as other fundamentals (debt levels and deficits) were fine at that time. Russia was notinsolvent based on debt levels and external flow imbalances (current account deficits) but its failure tomake flow fiscal adjustments and large capital flight led to default in 1998. 19 Node 13 <strong>of</strong> unsustainabledebt path with possible illiquidity included Argentina (2001), which did default, Indonesia (1997), whichdefaulted on many private external debts, and Thailand (1997) that did not default on most external debt(only on some private claims), but had severe debt servicing problems in its private financial and corporatesectors. There<strong>for</strong>e, these are cases <strong>of</strong> outright insolvency or high illiquidity that made the debt path notsustainable/financeable.VII. Regression Analysis18 Given the lack <strong>of</strong> good data on primary balances, our classification does not capture well cases, such as Turkey,where the country would be deemed as insolvent based on a debt level criteria, but may be solvent as it is running avery large primary surplus (above 5 percent <strong>of</strong> GDP currently), that is stabilizing and reducing such high debt level.19 In principle, we may have included a variable measuring “contagion” effects from crises elsewhere. In practice,however, implementation in our context is problematic, since contagion typically occurs within a year, and weemploy (one period lagged) data at annual frequency.


- 23 -A natural question is whether the non-linear interactions and threshold effects found in the tree arestatistically significant. We apply standard econometric techniques on our data set. 20 We proceed asfollows: we construct a set <strong>of</strong> dummy variables corresponding to the inequalities that define the terminalnodes in our classification tree (variables “node 1-14”, see Table 3), and estimate a logit model with ourcrisis indicator as dependent variables and the node dummies as dependent variables. We report the resultsin Table 4. [insert table 4 here ] Clearly, the node dummies corresponding to a “perfect classification”(nodes 1,3,4,6,8,9 in Figure 1) are dropped, as they perfectly predict the outcome. Also, not surprisingly,the “small” nodes containing only a handful <strong>of</strong> observations (nodes 5,10,12) are not found statisticallysignificant. Conversely, the large nodes have statistically significant coefficients, generally with theexpected sign. If a country displays “safe fundamentals”, as defined by Node 3 and Node 11 (see Table3), the probability <strong>of</strong> experiencing a crisis next year is significantly reduced; conversely, belonging toNode 14 <strong>of</strong> high inflation and debt significantly raises the probability <strong>of</strong> a crisis; the “liquidity crisis” node7 is the only dummy variable which is appears with the wrong sign, although it is only marginallysignificant.VIII. RobustnessAs we discussed in the Methodology section, robustness is a serious issue <strong>for</strong> classification trees. Here, webriefly describe and apply a solution recently developed by Breiman (2001), one <strong>of</strong> CART developers. The“Random Forests” (RF) algorithm 21 is a generalization <strong>of</strong> CART based on a collection <strong>of</strong> hundreds(possibly thousands) <strong>of</strong> trees. The idea is to build many trees using different combinations <strong>of</strong> variables anddata (sub) samples, in order to make sure that the inclusion or exclusion <strong>of</strong> variables and data-points doesnot affect the classification. Here is how it works. RF obtains the final classification (crisis vs. non-crisis)by “majority rule”, that is, by aggregating the responses <strong>of</strong> many trees. At each run, a bootstrap procedureestimates each tree by randomizing over two dimensions: a sub-sample <strong>of</strong> data and a (small) subset <strong>of</strong> all20 Ideally, one should per<strong>for</strong>m such a test on a “fresh” set <strong>of</strong> data.21 We are not aware <strong>of</strong> other applications <strong>of</strong> the Random Forest algorithm to international macro/finance issues.


- 24 -the variables. By picking a random subset <strong>of</strong> the variables, any single variable is analyzed in a differentcombination with others, thus taking care <strong>of</strong> the problem <strong>of</strong> “masking” discussed in the Methodologysection. Moreover, by randomizing simultaneously over the sub-samples, each tree analyzes only a portion<strong>of</strong> data at a time. This process, called “slow learning,” highlights different aspects <strong>of</strong> the data set andreduces the risk <strong>of</strong> possibly drawing wrong conclusions “too fast” (see Friedman, 2001) from a uniquesample. If a pattern genuinely exists in the portion <strong>of</strong> data analyzed, the RF algorithm will detect itrepeatedly in different trees; conversely, it will wash out any accidental pattern in the process <strong>of</strong> averagingthe results. The drawback <strong>of</strong> the algorithm is that it does not allow the researcher to recover a “single” treewith its thresholds and interactions, since it relies upon the aggregation <strong>of</strong> the classifications that producedby many different trees. For our purposes, however, Random Forest provides an interesting benchmark <strong>for</strong>assessing the robustness <strong>of</strong> CART’ selection <strong>of</strong> variables. 22 Table 5 [insert Table 5 here] reports the Ginimeasure <strong>of</strong> the variable importance <strong>for</strong> CART and RF 23 , normalized on a 0 (lowest importance)-100(highest) scale. Overall, this analysis confirms the robustness <strong>of</strong> our CART results. The solvencyindicators (total external debt over GDP, public external debt over revenue), the liquidity indicator (inparticular short-term debt over reserves), the exchange rate volatility and overvaluation, inflation andgrowth, all feature prominently in RF’s as well as in CART’s selection <strong>of</strong> most important variables.Interestingly, RF ranks the exchange rate variables even higher than CART. Also, political economyvariables do a better job in the RF than in the CART tree (see the indexes <strong>of</strong> Political Constraint). Theproximity to presidential elections is confirmed as a relevant predictor. Overall, the coefficient <strong>of</strong>correlation between the measures <strong>of</strong> importance in CART and RF is 0.61, and the rank correlation is 0.77.22 Each tree in Random Forest was built by selecting 3 variables at random, and 500 trees were built using <strong>of</strong> tenth <strong>of</strong>the observations at each run.23 In practice, every variable is considered at every split, in each tree, and the Gini index provides a measure <strong>of</strong> the“purity” <strong>of</strong> the sub-nodes generated by the variable split, see the discussion in Section IV, and Breiman et al (1984).


- 25 -IX. PredictionNext, we discuss the prediction accuracy <strong>of</strong> our model. Table 6 [insert Table 6 here] reports the predictiveaccuracy <strong>of</strong> the tree <strong>for</strong> both crisis “states” and crisis “entries”, being respectively defined as crisespreceded by a crisis, and by a non-crisis. The purpose <strong>of</strong> this distinction is to facilitate the comparisonbetween our results and those obtained in the Early Warnings literature. Many empirical models focusonly on entries, see <strong>for</strong> example Chamon, Manasse and Prati (2006). The motivation is that factors thatlead to a crisis may be different (or have a different impact) from factors that keep the country in the crisisstate. Neglecting all within-crisis observations, however, does not make an efficient use <strong>of</strong> availablein<strong>for</strong>mation. In fact, Manasse, Roubini, Schimmelpfennig, (2003) find that the same variables explain bothcrisis entry and persistence, typically with statistically equal coefficients. In our estimation, we have keptall crisis states in<strong>for</strong>mation, <strong>for</strong> two additional reasons. First, because CART treats each observationindependently, so that Argentina 1994 and Argentina 1995 are effectively different “individuals”. Second,because CART works best with large data sets.Comparing our results with those in the EWS literature is not immediate, since the definitions <strong>of</strong> crisis, thedata frequency (monthly, yearly), the sample coverage, and the <strong>for</strong>ecast horizon (one, two years ahead)differ dramatically across studies. The most successful EWS models are able to correctly predict 60-70percent <strong>of</strong> (currency) crises (<strong>for</strong> example, Berg and Patillo, 1999), but they differ widely in their falsealarms, and in their in and out-<strong>of</strong> sample predictions (which typically per<strong>for</strong>m much worse) . Clearly, onecould easily predict as many crises as wished, simply by accepting a very large number <strong>of</strong> false alarms (ina logit model one would choose a sufficiently low crisis probability threshold). Berg and Patillo, 1999,find that in this respect the models they analyze <strong>for</strong> currency crises yield disappointing results: between 63and 54 percent <strong>of</strong> false alarms, with a crisis probability thresholds <strong>of</strong> 25 percent 24 .24 These results were found <strong>for</strong> out-<strong>of</strong>-sample predictions


- 26 -Table 6 presents the results. The first two columns give the in-sample predictive accuracy <strong>for</strong> the modelbased on the full sample, both <strong>for</strong> the entire set <strong>of</strong> crises and when applied only to the post-1989 crises.The tree per<strong>for</strong>ms quite well when compared with the standard EWS models. It correctly predicts oneyear ahead 82.9 percent <strong>of</strong> the states, and does only slightly worse <strong>for</strong> the post 1990 period; it does betterwith crisis states, <strong>of</strong> which it correctly predicts 93.9 percent, than with tranquil states (not shown:79.9 percent). The prediction success <strong>for</strong> “entries” is comparably good (88.9 and 85 percent in full andpost 1989 samples); moreover, these results do not come at the expenses <strong>of</strong> considerable false alarms: <strong>for</strong>the post 1989, only 15 percent <strong>of</strong> tranquil states are incorrectly predicted as crises (and 18,5 percent <strong>for</strong> theentire sample). These results also compare quite favorably with those obtained by Manasse, Roubini andSchimmelpfennig (2003). The logit model estimated in this paper, <strong>for</strong> a slightly higher threshold, hasfewer false alarms (5.7 percent) but also correctly predicts far fewer entries (57.1 percent).“Out” <strong>of</strong> sample predictions are obtained as follows. First, we build a tree is built using only data only upto 1989. Here we <strong>for</strong>ce CART to use the same set <strong>of</strong> 10 variables found in the optimal tree based on theentire sample 25 . Predictions <strong>for</strong> the post- 1989 data are then obtained using this tree.The out <strong>of</strong> sample prediction accuracy is considerably worse than in-sample. About 57 percent <strong>of</strong> defaultstates and 35 percent <strong>of</strong> post-1990 entries are correctly predicted from the tree based on pre-1990observations. In particular, the model fails to predict many <strong>of</strong> the important crises <strong>of</strong> the 1990’s: the Asiancrisis, Russia, Brazil and Turkey (2001) are not predicted 26 . Note however, that we are requiring ourmodel to predict as far as 11 years ahead, quite a <strong>for</strong>midable task. Moreover, false alarms (incorrectly25 The resulting tree and threshold is only marginally different from the optimal one based on the full data set. Details areavailable on request. Alternatively , we could have let CART choose a potentially new set <strong>of</strong> variables in the estimation .Doing so, however, does not improve the predictive accuracy.26 The following crisis entries are missed: Indonesia (1997, 2002), Korea (1997), Thailand (1997), Argentina (1995),Brazil (1998), Mexico (1995), Russia(1998),Tunisia(1991),Turkey(2001), Ukraine (1998), Venezuela(1995).Conversely Algeria (1991), Argentina(2001), Ecuador(1999), Pakistan (1998), South Africa(1993), Uruguay(1990)and Venezuela(1990) are correctly predicted.


- 27 -called entries) are also extremely low, which suggest that experimenting with different thresholds mayimprove the model predictive power to some extent. 27 .X. Rolling Trees and Changes in Variables and ThresholdsThe unsatisfactory out-<strong>of</strong>-sample per<strong>for</strong>mance <strong>of</strong> the model is likely to be due to the changingmacroeconomic configurations <strong>of</strong> crisis episodes across different decades. Within our framework, thisshould be reflected either by a change in the set <strong>of</strong> explanatory variables and/or by a change in theircritical thresholds over time. We examine both issues in by running “rolling trees”.The first exercise asks the question: has the set <strong>of</strong> optimal crisis-predictors in emerging markets changedover time? We run unrestricted rolling trees, by letting CART choose the best set <strong>of</strong> predictors <strong>for</strong> eachrun, starting from a 1970-1989 sample, and then adding one year at a time, until the full sample is covered.For each run, we calculate the variables’ “importance” measure previously discussed, and plot thenormalized values (from 0 to 100) in Figure 2. Note that different trees <strong>of</strong>ten pick different variables <strong>for</strong>very similar indicators (<strong>for</strong> example, <strong>for</strong> solvency, Total (= public + private) External <strong>Debt</strong>/GDP issometimes replaced by Total External <strong>Debt</strong>/Exports or by Public External <strong>Debt</strong>/GDP). 28 In the Figure wepick the variable with the highest importance measure <strong>for</strong> each indicator (see the note to the Figure)Figure 2 shows that the size <strong>of</strong> External <strong>Debt</strong> is consistently the best predictor <strong>of</strong> sovereign debt crises inemerging markets. Measures <strong>of</strong> liquidity (Short Term <strong>Debt</strong>), the second most important predictor up to theearly 1990s, have become temporarily less important in the 1995-98 period, when fiscal solvencyconsiderations (Public External <strong>Debt</strong>/Revenue), political uncertainty (Years to Presidential elections) andinflation have gained predictive power. 29 This makes sense, since in addition to the Asian crises, this27 When the estimation sample is extended to 1996, the predictive accuracy <strong>of</strong> the model does not improvesignificantly, possibly confirming a sort <strong>of</strong> “structural break” in the latter crises We have not pursued this attemptsince we wanted to keep the model parameters as closed as possible to those used in the in-sample exercise.28 Recall that a variable may score high even if it does not appear in the tree (<strong>for</strong> example, a variable may be a“second-best” split at each stage)29 Note however that the Financial Requirement indicator (Short term external debt plus current account deficit overReserves has somewhat gained in importance) .


- 28 -period witnesses, among others, crises in Algeria 1995, 1996, Argentina 1995, Ecuador 1995, Mexico1995, Pakistan 1998 , Peru 1995, 1996 , 1997, Russia 1998, Ukraine 1998 and Venezuela 1995,1996. Other indicators show a secondary and stable role.Next we ask the question: have the critical thresholds changed over time in emerging markets? Thisquestion has no obvious answer in our tree framework, since following the top split, classification treescalculate conditional thresholds: the same variable may display a different thresholds, depending on thevariable’s place in the tree’s sequence. For this reason we adopt the following strategy: we run rollingtrees as be<strong>for</strong>e, but this time keeping only the variables that appear in the “final” (full sample) tree in eachrun, as we did in the out-<strong>of</strong>.-sample exercise. For each different tree, we report the (unconditional)threshold value that each variable would have taken, had it been selected in the top split.The thresholds’ evolution <strong>for</strong> the main variables s are shown in Figure 3. The critical value <strong>for</strong> thesolvency indicator, Total External <strong>Debt</strong>/GDP, slowly creeps up from 0.44 to 0.5 , possibly indicatingmore resilience to large debts in more recent. episodes. By contrast, the threshold <strong>for</strong> Short-Term<strong>Debt</strong>/Reserves slightly falls from 83 to 73 percent since 1994, suggesting that although liquidity indicatorsmay have lost some predictive power in the late 1990s (Figure 2), the vulnerability <strong>of</strong> emerging markets toliquidity shocks may have risen somewhat. Interestingly, the temporary fall in the threshold <strong>of</strong> PublicExternal <strong>Debt</strong>/Revenue, from 2 to 1.7 in the second half <strong>of</strong> the 1990s, confirms the larger role <strong>of</strong> publicsector vulnerability in the crises <strong>of</strong> this period. Finally, it is interesting to note that, possibly as a reflection<strong>of</strong> larger capital market integration in more recent years, the critical value <strong>of</strong> an over-appreciatedexchange rate (not shown) falls sharply, from 100 to 34 percent.In summary, the rolling tree analysis suggests that the nature <strong>of</strong> sovereign debt crises changed somewhatin the nineties, with fiscal solvency and political economy issues gaining more relevance, and withemerging economies showing heightened vulnerability to (possibly less frequent) liquidity shocks and tocurrency misalignments, possibly reflecting their larger integration in the international capital markets.The combination <strong>of</strong> changes in the ranking <strong>of</strong> the explanatory variables and in their critical thresholdshelps explaining why the more recent crises are difficult to predict on the basis <strong>of</strong> the earlier ones.


- 29 -XI. ConclusionsIn this paper we have applied a new statistical methodology to the question <strong>of</strong> understanding sovereigndebt crises, both in terms <strong>of</strong> fundamentals that lead to a crisis, and <strong>of</strong> the factors that allow us to predictsuch crises. This technique allows us to derive endogenously the most important factors associated tovulnerability <strong>of</strong> sovereign debt, and the thresholds that signal greater risk <strong>of</strong> a crisis.We find that most debt crises can be classified into three types: i) episodes <strong>of</strong> insolvency (high debt andhigh inflation) or debt unsustainability due to high debt and illiquidity; ii) episodes <strong>of</strong> illiquidity, wherenear default is driven by large stocks <strong>of</strong> short-term liabilities relative to <strong>for</strong>eign reserves; and iii) episodes<strong>of</strong> macro and exchange rate weaknesses (large overvaluation and negative growth shocks). Conversely, arelatively “risk-free” country type is described by a handful <strong>of</strong> economic characteristics: low total externaldebt relative to ability to pay, low short-term debt over <strong>for</strong>eign reserves, low public external debt overfiscal revenue, and an exchange rate that is not excessively overvalued. Political instability and tightmonetary conditions in international financial markets aggravate liquidity problems. The approachsuggests that unconditional thresholds—<strong>for</strong> example, looking at debt to output ratios in isolation—are <strong>of</strong>little value per se <strong>for</strong> assessing the probability <strong>of</strong> default; it is the particular mix <strong>of</strong> different types <strong>of</strong>vulnerability that may lead to a sovereign debt crisis.The in-sample predictive power <strong>of</strong> the tree approach is quite good with very few crises missed; the modelpredictive accuracy does not suffer when applied to the post-1990 experience. However, despite very fewfalse alarms, the out-<strong>of</strong> sample predictive power is not satisfactory: many <strong>of</strong> the crises <strong>of</strong> 1990s could notbe predicted based on previous decades in<strong>for</strong>mation, possibly confirming the different nature <strong>of</strong> the lattercrises. The “rolling tree” analysis suggests indeed that the nature <strong>of</strong> sovereign debt crises changedsomewhat in the nineties, with fiscal solvency and political economy issues gaining more relevance, andvulnerability to liquidity shocks and currency over-appreciation becoming larger.Our approach allows to derive rules <strong>of</strong> thumb which may be useful to predict crises early on (an “earlywarnings signal” model), and to derive policy adjustment paths, which may reduce the likelihood <strong>of</strong> a


- 30 -crisis <strong>for</strong> countries that may be entering in a danger zone. Thus, ideally, this tool can be used <strong>for</strong>surveillance, crisis prevention, and crisis resolution.A number <strong>of</strong> possible extensions come to mind, some <strong>of</strong> which we have tried. First, following the idea thata “history” <strong>of</strong> past defaults may bear on the credibility <strong>of</strong> a sovereign and thus affect the crisis probability(as suggested by Reinhart and others, 2003, in their “debt intolerance” hypothesis), we constructed a newvariable taking on incremental values each time a new default episode occurred. Our indicator <strong>of</strong> (bad)“reputation” did not affect the classification tree. Experimenting with contagion effect is also possible. Butthe use <strong>of</strong> annual data preclude using same-year in<strong>for</strong>mation <strong>for</strong> predicting crises.


- 31 -ReferencesArias, Andres F., 2004, “Comments on the paper ‘Assessing Fiscal Sustainability;’ by Enrique G.Mendoza and Pedro Marcelo Oviedo.” Available via the Internethttp://www.iadb.org/res/centralbanks/publications/cbm35_199.pptBerg, Andrew, and Catherine Pattillo, 1999, “Are Currency <strong>Crises</strong> Predictable? A Test,” StaffPapers, Vol. 46 (June), No. 2, pp. 107–38.Breiman, L. (2001). “Random Forest”, Machine Learning,http://oz.berkeley.edu/users/breiman/random<strong>for</strong>est2001.pdfBreiman, Leo, Jerome H. Friedman, Richard A. Olshen, and Charles J. Stone, 1984, Classification andRegression Trees, (London: Chapman & Hall).Beers, David T., and Ashok Bhatia, 1999, “<strong>Sovereign</strong> Defaults: History,” in Standard & Poor’s CreditWeek, December 22.______, and John Chambers, 2002, “<strong>Sovereign</strong> Defaults: Moving Higher Again in 2003?,” Standard &Poor’s (September). Available via the Internet http://www.emta.org/keyper/sov2003.pdfBeim, David O., and Charles W. Calomiris, 2001, Emerging Financial Markets, (New York: McGraw-Hill/Irwin).Cantor, Richard, and Frank Packer, 1996, “Determinants and Impact <strong>of</strong> <strong>Sovereign</strong> Credit Ratings,”Federal Reserve Bank <strong>of</strong> New York Policy Review (October), pp. 37–52.Catão, Luis, and Bennett Sutton, 2002, “<strong>Sovereign</strong> Defaults: The Role <strong>of</strong> Volatility,” IMF Working Paper02/149 (Washington: International Monetary Fund).Corsetti, Giancarlo, Bernardo Guimarães, and Nouriel Roubini, 2003, “The Trade<strong>of</strong>f Between anInternational Lender <strong>of</strong> Last Resort to Deal with Liquidity Crisis and Moral Hazard Distortions: AModel <strong>of</strong> the IMF’s Catalytic Finance Approach,” IMF Seminar Series Research Paper 03/149(Washington: International Monetary Fund).


- 32 -Cottarelli, Carlo, and Curzio Giannini, 2002, “Bedfellows, Hostages, or Perfect Strangers? Global CapitalMarkets and the Catalytic Effect <strong>of</strong> IMF Crisis Lending,” IMF Working Paper 02/193(Washington: International Monetary Fund).Dell’Arriccia, Giovanni, Isabel Schnabel, and Jeromin Zettelmeyer, 2002, “Moral Hazard andInternational Crisis Lending: A Test,” IMF Working Paper 02/181 (Washington: InternationalMonetary Fund).Detragiache, Enrica, and Antonio Spilimbergo, 2001, “<strong>Crises</strong> and Liquidity: Evidence and Interpretation,”IMF Working Paper 01/2 (Washington: International Monetary Fund).Eaton, Jonathan, and Raquel Fernandez, 1995, “<strong>Sovereign</strong> <strong>Debt</strong>,” NBER Working Paper 5131, Prepared<strong>for</strong> Handbook <strong>of</strong> International Economics.Frankel, Jeffrey, and Shang-Jin Wei, 2004, “Managing Macroeconomic <strong>Crises</strong>: Policy Lessons,” NBERWorking Paper 10907, November. Available via the Internethttp://ksghome.harvard.edu/~jfrankel/Managing%20Macroeconomic%20<strong>Crises</strong>%20Policy%20Lessons.pdfGhosh, Swati, and Atish Ghosh, 2002, “Structural Vulnerabilities and Currency <strong>Crises</strong>,” IMF WorkingPaper 02/9 (Washington: International Monetary Fund).Haque, Nadeem U., Mark Nelson, and Donald J. Mathieson, 1998, “The Relative Importance <strong>of</strong> Politicaland Economic Variables in Creditworthiness Ratings,” IMF Working Paper 98/46 (Washington:International Monetary Fund).Hemming, Richard, and Nigel Chalk, 2000, “Assessing Fiscal Sustainability in Theory and Practice,” IMFWorking Paper 00/81 (Washington: International Monetary Fund).______, and Murray Petrie, 2002, “A Framework <strong>for</strong> Assessing Fiscal Vulnerability,” IMF Working Paper00/52 (Washington: International Monetary Fund).Henisz, W. J., (2002), “The Institutional Environment <strong>for</strong> Infrastructure Investment,” Industrial andCorporate Change, Vol.11, Number 2, pp. 355–89.


- 33 -Jeanne, Olivier, 2000, “<strong>Debt</strong> Maturity and the Global Financial Architecture,” CEPR Discussion Paper2520 (London: Centre <strong>for</strong> Economic Policy Research).Lane, Timothy, and Steven Phillips, 2001, “Does IMF Financing Result in Moral Hazard?” IMF WorkingPaper 00/168 (Washington: International Monetary Fund).Larrain, Guillermo, Helmut Reisen, and Julia von Maltzan, 1997, “Emerging Market Risk and <strong>Sovereign</strong>Credit Ratings,” OECD Development Centre Technical Paper 124 (Paris: Organization <strong>for</strong>Economic Co-operation and Development).Lee, Suk Hun, 1993, “Are the credit ratings assigned by bankers based on the willingness <strong>of</strong> LDCborrowers to repay?,” Journal <strong>of</strong> Development Economics, Vol. 40, pp. 349–59.Manasse, Paolo, Nouriel Roubini, and Axel Schimmelpfennig, 2003,“Predicting <strong>Sovereign</strong> <strong>Debt</strong> <strong>Crises</strong>,”IMF Working Paper 03/221 (Washington: International Monetary Fund).Mody, Ashoka, and Diego Saravia, 2003, “Catalyzing Private Capital Flows: Do IMF-Supported ProgramsWork as Commitment Devices?” (unpublished; Washington: International Monetary Fund).Obstfeld, Maurice, and Kenneth Rog<strong>of</strong>f, 1996, “Foundations <strong>of</strong> International Macroeconomics,” MITPress (Cambridge, Massachusetts).Reinhart, Carmen M., 2002, “Default, Currency <strong>Crises</strong> and <strong>Sovereign</strong> Credit Ratings,” NBER WorkingPaper 8738. Also in The World Bank Economic Review, Vol. 16, No. 2, pp. 151–70.Reinhart, Carmen M., Kenneth S. Rog<strong>of</strong>f, and Miguel A. Savastano, 2003,“<strong>Debt</strong> Intolerance,” NBERWorking Paper 9908 (Cambridge, Massachusetts: National Bureau <strong>of</strong> Economic Research).Rojas-Suarez, L., 2001, Rating Banks in Emerging Markets, (Washington: Institute <strong>for</strong> InternationalEconomics).Roubini, Nouriel, 2001, “<strong>Debt</strong> Sustainability: How to Assess Whether a Country is Insolvent,”unpublished (New York: New York University) (December).


- 34 -Roubini, Nouriel, and Brad Setser, 2004, Bailouts or Bail-ins? Responding to Financial <strong>Crises</strong> inEmerging Economies, (Washington: Institute <strong>for</strong> International Economics).


- 35 -Tables and FiguresTable 1. Countries and Default Episodes in the Full SampleNumber Average Years<strong>of</strong> Crisis Length in CrisisCrisis Episodes (starred are added byIMF loans)Algeria 1 6.0 6 1991–97Argentina 3 5.0 15 1982–94, 1995*,1996, 2001,2002Bolivia 2 6.5 13 1980–85, 1986–94Brazil 3 5.3 16 1983–95, 1998*,99*,2000, 2001*,2002*Chile 1 8.0 8 1983–91China 0 … 0Colombia 0 … 0Costa Rica 1 10. 10 1981–91Cyprus 0 … 0Czech Republic 1/ 0 … 0Dominican Republic 1 22. 22 1981–Ecuador 2 8.0 16 1982–96, 1999–01Egypt 1 1.0 1 1984–85El Salvador 1 16. 16 1981–97Estonia 1/ 0 … 0Guatemala 1 1.0 1 1986–87Hungary 1/ 0 … 0India 0 … 0Indonesia 2 2.5 5 1997*–01, 2002Israel 0 … 0Jamaica 3 4.7 14 1978–80, 1981–86, 1987–94Jordan 1 5.0 5 1989–94Kazakhstan 1/ 0 … 0Korea, Rep. <strong>of</strong> 2 2.0 4 1980*,81*,82, 1997*,98*,99Latvia 1/ 0 … 0Lithuania 1/ 0 … 0Malaysia 0 … 0Mexico 2 5.0 10 1982–91, 1995*, 96Morocco 2 3.0 6 1983–84, 1986–91Oman 0 … 0Pakistan 1 2.0 2 1998–00Panama 1 14. 14 1983–97Paraguay 1 7.0 7 1986–93Peru 3 6.3 19 1976–77, 1978,79*–81, 1983–98Philippines 1 10. 10 1983–93Poland 1/ 0 … 0Romania 1/ 0 … 0Russia 1/ 1 3.0 3 1998–01Slovak Republic 1/ 0 … 0South Africa 4 1.8 7 1976*,77*,78, 1985–88, 1989–90, 1993–Thailand 2 1.0 2 1981*,82, 1997*, 98Trinidad and Tobago 1 2.0 2 1988–90Tunisia 1 1.0 1 1991*, 92Turkey 2 3.5 7 1978,79,80*,81*,82, 83, 2000*,01*,02Ukraine 1/ 1 3.0 3 1998–01Uruguay 3 2.0 6 1983–86, 1987–88, 1990–92Venezuela 3 3.3 10 1983–89, 1990–91, 1995–98Total 54 5.5 261Sources: IMF, Standard & Poor’s, World Bank, and authors’ calculations.. A star indicates according to IMF loan


- 36 -Table 2. Mean <strong>of</strong> Variables Used in the AnalysisCurrent year All No crisis No crisis Crisis Crisis No. <strong>of</strong>Next year All No crisis Crisis Crisis No crisis Obs.Total external debt over GDP 45.5 37.0 54.7 71.4 63.7 1,051Total external debt over exports 290.7 239.3 359.3 455.9 350.2 1,053Total debt service over GDPTotal debt service over reservesShort-term external debt (RM, over GDP) 10.9 9.4 15.0 15.1 15.7 993Short-term external debt (RM, over reserves) 1.7 1.2 2.9 2.9 2.2 948Interest on short-term external debt over GDP 0.5 0.5 0.8 0.6 0.7 754Interest on short-term external debt over reserves 0.1 0.1 0.2 0.1 0.1 754<strong>Debt</strong> service on short-term external debt over GDP 5.3 4.8 6.9 6.4 7.1 1,050<strong>Debt</strong> service on short-term external debt over reserves 0.8 0.7 1.5 1.2 0.9 1,050Public external debt over GDP 32.2 25.5 36.4 53.0 46.5 1,051Public external debt over revenue 1.7 1.3 1.9 3.0 2.3 827Consolidated central government debt 47.5 46.4 38.2 57.3 54.0 462Overvaluation 0.0 0.0 0.0 -0.1 0.0 799Current account balance -2.7 -2.7 -4.3 -2.6 -1.3 1,231Reserves growth 19.1 20.8 -5.4 17.8 22.9 1,155Export growth 12.0 13.8 4.9 6.4 7.8 1,270M2 over reserves 5.6 5.3 7.9 6.2 6.2 1,185Financing requirement 1.6 1.3 3.0 2.4 1.4 986LIBOR 9.7 9.5 10.5 10.5 9.4 1,229U.S. treasury bill rate 6.4 6.3 7.8 6.9 6.3 1,229Inflation 54.6 17.5 241.1 169.6 84.9 1,273Nominal GDP growth 55.9 22.8 249.4 148.6 96.0 1,276Real GDP growth 4.1 4.8 1.8 2.1 2.2 1,276REER growth 121.5 124.3 139.7 111.0 109.4 937Import growth 10.0 12.3 5.3 4.8 6.9 902FDI over GDP 1.7 1.9 1.1 1.0 1.5 1,024Openness 71.2 71.3 64.1 72.1 72.5 903Overall balance -4.3 -4.4 -6.3 -3.8 -4.1 1,013Primary balance 0.6 0.3 -0.9 2.0 1.5 616Primary gap 6.6 5.7 23.1 59.0 -15.0 122G.Gov revenue over GDP 24.3 25.4 22.7 20.1 24.2 1,013Overall Balance/GDP variability 1.92 -2.53 -1.77 19.42 -1.70 824Inflation variability 120.43 10.80 125.37 549.81 83.03 1,082Exchange rate variability 2/ 57.8 56.38 45.62 39.87 159.9 1,059Real Exchange Rate Variability 2.87 1.49 3.76 6.34 3.39 753Terms <strong>of</strong> trade variability 2.24 2.28 2.32 2.23 1.33 1,088Index <strong>of</strong> political rights 3.52 3.63 3.44 3.18 3.20 1,130Index <strong>of</strong> civil liberties 3.77 3.85 3.75 3.44 3.65 1,130Index <strong>of</strong> freedom status 1.21 1.16 1.21 1.39 1.39 1,130Index <strong>of</strong> political constraints III 0.24 0.24 0.23 0.26 0.31 1,229Index <strong>of</strong> political constraints V 0.35 0.33 0.35 0.37 0.43 1,229Year <strong>of</strong> parliamentary elections 0.19 0.18 0.24 0.23 0.20 1,276Year <strong>of</strong> presidential elections 0.09 0.06 0.11 0.18 0.15 1,276Year <strong>of</strong> parliamentary or presidential elections 0.21 0.20 0.24 0.27 0.24 1,276Years to next parliamentary elections 2.73 2.94 2.67 2.06 2.16 1,071Years to next presidential elections 5.02 6.47 3.15 2.18 2.39 660Years since parliamentary wlections 2.02 1.81 2.47 2.60 2.00 901Years since presidential elections 2.19 2.15 2.83 2.24 1.60 418Electoral system 1/ 1 1 1 1 1 1,2101/ Mode <strong>of</strong> electoral system.2/ Excludes Turkey. Sources: IMF, Standard & Poor’s, World Bank, and authors calculations


Table 3. Classification TableColumn 1/TerminalNode2 3 4 5 6 7 8 9Rules and Critical ThresholdsA. Relatively SafeTotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res ≤ 1.3,1 PublicExt<strong>Debt</strong>/Revenue ≤ 2.1, GDPGrowth ≤ -5.5,EchROverV ≤ 48N. Crisis % <strong>of</strong> % <strong>of</strong> Relative ClassifiedEpisodes Likelihood <strong>Crises</strong> NonCrisis Safety Crisis (1)<strong>Crises</strong>In Node In Node In Node In Node <strong>of</strong> Node NonCrisis(0) (entry)13 0.0 0.0 1.3 1.26 0468911TotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res ≤ 1.3,PublicExt<strong>Debt</strong>/Revenue > 2.1, Inflation ≤ 10.7TotExt<strong>Debt</strong>/GDP ≤ 19.1, ShortT<strong>Debt</strong>/Res > 1.3,YearNextPresElect ≤ 5.5TotExt<strong>Debt</strong>/GDP ≤ 49.7, TotExt<strong>Debt</strong>/GDP > 19.1,ShortT<strong>Debt</strong>/Res > 1.3, YearNextPresElect ≤ 5.5,EchRVolatil > 27.9TotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res > 1.3,YearNextPresElect > 5.5, USTresBill ≤ 9.7TotExt<strong>Debt</strong>/GDP > 49.7, Inflation ≤ 10.5,ExtFinReqs/Reserve ≤ 1.4,PublicExt<strong>Debt</strong>/Revenue ≤ 3.113 0.0 0.0 1.3 1.26 023 0.0 0.0 2.3 1.26 022 0.0 0.0 2.2 1.26 051 0.0 0.0 5.0 1.26 098 2.0 0.8 9.5 1.23 0 Algeria (1991)3B. Liquidity RiskTotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res ≤ 1.3,PublicExt<strong>Debt</strong>/Revenue ≤ 2.1, GDPGrowth > -5.5607 2.3 5.4 58.4 1.23 0Venezuela (1983), South Africa (1976), Russia (1998),Argentina (1995), Thailand (1981)- 37 -10TotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res > 1.3,YearNextPresElect > 5.5, USTresBill > 9.710 40.0 1.5 0.6 0.75 17TotExt<strong>Debt</strong>/GDP ≤ 49.7, TotExt<strong>Debt</strong>/GDP > 19.1,ShortT<strong>Debt</strong>/Res > 1.3, YearNextPresElect ≤ 5.5,EchRVolatil ≤ 27.8C. Solvency RiskTotExt<strong>Debt</strong>/GDP > 49.7, Inflation ≤ 10.5,12 ExtFinReqs/Reserve ≤ 1.4,PublicExt<strong>Debt</strong>/Revenue> 3.113TotExt<strong>Debt</strong>/GDP > 49.7, Inflation ≤ 10.5,ExtFinReqs/Reserve > 1.4130 41.5 20.7 7.5 0.73 110 40.0 1.5 0.6 0.75 179 46.8 14.2 4.1 0.67 1Pakistan (1998), Mexico (1982, 1995), Argentina (1982, 1989,1993), Peru (1976, 1983), Brazil (1998, 2001), Jamaica (1978),Turkey (1978), Korea (1997), Trinidad & Tobago (1988), Ukraine(1988), Dominican Republic (1981)Chile (1993), Tunisia (1991), Panama (1983), Thailand (1997),Indonesia (1997), Philippines (1983), Argentina (2001), Jordan(1989), Morocco (1986), Costa Rica (1981)5TotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res ≤ 1.3,PublicExt<strong>Debt</strong>/Revenue > 2.1, Inflation > 10.620 55.0 4.2 0.9 0.57 1 Guatemala (1986), Paraguay (1986)14 TotExt<strong>Debt</strong>/GDP > 49.7, Inflation > 10.5 196 66.9 50.2 6.4 0.42 1D. Fiscal, Exchange rate and MacrorisksTotExt<strong>Debt</strong>/GDP ≤ 49.7, ShortT<strong>Debt</strong>/Res ≤ 1.3,2 PublicExt<strong>Debt</strong>/Revenue ≤ 2.1, GDPGrowth ≤ -5.5,EchROverV > 48Total 1276 100.0 100.0Source: Author’s Calculations.Jamaica (1981, 1987), Egypt (1984), Bolivia (1986), Peru (1978),Ecuador (1982, 1999), Uruguay (1987), Indonesia (2002),Bolivia (1980), Morocco (1983), Turkey (2000), South Africa(1985), Uruguay (1990), Brazil (1983), Venezuela (1990, 1995)4 100.0 1.5 0.0 0.00 1 El Salvador (1981), Uruguay (1983)


- 38 -Table 4. Logit estimatesNumber <strong>of</strong> obs = 1163Log likelihood = -379.97026__________________________________________________________Crisis Coef. Std. Err. z P>|z| [95% Conf. Interval]___________________________________________________________node3 -3.746137 .2703977 -13.85 0.000 -4.276107 -3.216167node5 .2006707 .4494666 0.45 0.655 -.6802676 1.081609node7 -.3417493 .1779787 -1.92 0.055 -.6905811 .0070825node10 -.4054651 .6454972 -0.63 0.530 -1.670616 .8596862node11 -3.871201 .7144343 -5.42 0.000 -5.271466 -2.470936node12 -.4054651 .6454972 -0.63 0.530 -1.670616 .8596862node13 -.1267517 .2254696 -0.56 0.574 -.568664 .3151606node14 .7008101 .1517175 4.62 0.000 .4034492 .9981709


- 39 -Table 5. Variable Importance and Rank in Cart and Random ForestCARTRandom ForestVariable Importance Rank Importance RankTot Ext <strong>Debt</strong>/GDP 100,00 1 100,00 1Pub Ext <strong>Debt</strong>/GDP 88,35 2 94,49 4Tot Ext <strong>Debt</strong>/Exp 69,35 3 79,17 6Sh TermExt<strong>Debt</strong>/Res (rem mat) 38,87 4 79,63 5ShTermExt<strong>Debt</strong>/Res (orig mat) 34,55 5 63,09 9PubExt<strong>Debt</strong>/Res 33,71 6 94,63 3ShTermExt<strong>Debt</strong>/GDP(rem.mat) 29,85 7 42,92 11ExtFinReqs/Res 28,23 8 34,65 16GDP NominalGrowth 25,50 9 33,49 19Inflation 24,75 10 65,34 8ServExt<strong>Debt</strong>/GDP 24,40 11 38,01 12<strong>Debt</strong>ServExt<strong>Debt</strong>/Res 20,89 12 35,19 15M2/Res 19,73 13 26,61 21ExchRateOverv 15,56 14 33,71 18US TreasBill Rate 15,48 15 6,94 42YearsNextPresElec 14,30 16 17,69 31IntShTermExt<strong>Debt</strong>/Res 13,33 17 99,55 2ReserveGrowth(2) 11,95 18 15,62 32<strong>Debt</strong>ServExt<strong>Debt</strong>/GDP 11,61 19 22,49 27Exch Vol 11,43 20 20,61 29Inflation Vol 11,33 21 35,36 14Real GDP growth 8,56 22 23,23 26Revenue/GDP 8,35 23 22,4 28Openness 7,79 24 33,91 17PoliticalConstr III 7,00 25 27,77 20RealEchRateVol 7,57 25 67,58 7PoliticalConstr I 7,00 27 26,18 22ExpGrowth 5,98 29 13,3 35FDI/GDP 5,34 29 23,69 25LIBOR 4,45 30 9,99 38Civil Liberty Index 3,47 31 8,22 41M2 Growth 3,20 32 20,42 30PoliticalRights Index 2,95 33 9,72 39Electoral System 2,21 34 23,93 24IntShTermExt<strong>Debt</strong>/GDP 2,15 35 46,82 10GenGovBal/GDP Volat 1,33 36 15 33CurrAcc/GDP 0,97 37 10,19 37ReserveGrowth 0,34 38 25,53 23YearsSincePresElect 0,21 39 9,06 40YearsNextParlElect 0,10 40 5,08 45YearsSinceParlElect 0,00 41 4,24 46Terms<strong>of</strong>Trade Vol 0,00 42 14,65 34RealExchRateDep 0,00 43 37,07 13Public<strong>Debt</strong>/GDP 0,00 44 6,81 43PrimBalanceAdjReq 0,00 45 1,25 47PrimaryBalance/GDP 0,00 46 10,64 36PresElectionYear 0,00 47 1,02 48Pres&ParlElectYear 0,00 48 0,15 49ParlamElectYear 0,00 49 0 50FreedomStatusIndex 0,00 50 6,64 44Source: Authors Calculations


- 40 -Table 6. PredictionIn SampleFull Sample1990’s Onwards (Based on FullSample)Out <strong>of</strong> Sample1990’s Onwards (Based on 1970-1989)Observations 1276 556 556Number <strong>of</strong> crisisepisodes261 114 114Number <strong>of</strong> crisis entryepisodes 54 20 20Correctly calledepisodesCorrectly called defaultepisodesCorrectly called nondefault episodes(In percent )82,8 78,8 72,093,9 89,5 57,079,9 76,0 76,0Correctly called entriesinto default88,9 85,0 35,0Incorrectly called18,5 15,0 19,0entries into defaultSources: IMF, Standard & Poor’s, World Bank; and authors’ calculation.


- 41 -


Figure 1. The Empirical TreeTotal external debt> 50 percent <strong>of</strong> GDPNoYesShort-term external debtto reserves> 1.34Inflation> 10.47 percentNo Yes No YesPublic external debt Years to nextExternal financingNode 14to revenuepresidential electionrequirementObservations: 196> 2.15> 5.5> 1.44<strong>Crises</strong>: 66.8 percentNo Yes No Yes No YesReal GDP growth Inflation Total external debtUS T-bill ratePublic external debt Node 13> -5.45 percent > 10.67 percent > 19.1 percent <strong>of</strong> GDP> 9.72 percentto revenue> 3.1Observations: 79<strong>Crises</strong>: 46.8No Yes No Yes No Yes No Yes No Yes- 42 -Overvaluation> 48Node 3 Node 4 Node 5 Node 6 Exchange rate volatility Node 9 Node 10 Node 11 Node 12Observations: 607 Observations: 13 Observations: 20 Observations: 23 > 27.88 Observations: 51 Observations: 10 Observations: 98 Observations: 10<strong>Crises</strong>: 2.3 percent <strong>Crises</strong>: 0 percent <strong>Crises</strong>: 55 percent <strong>Crises</strong>: 0 percent <strong>Crises</strong>: 0 percent <strong>Crises</strong>: 40 percent <strong>Crises</strong>: 2 percent <strong>Crises</strong>: 40 percentNo Yes No YesNode 1 Node 2 Node 7 Node 8Observations: 13 Observations: 4 Observations: 130 Observations: 22<strong>Crises</strong>: 0 percent <strong>Crises</strong>:100 percent <strong>Crises</strong>: 41.5 percent <strong>Crises</strong>: 0 percentNot crisis-prone Crisis-prone Not crisis-prone Not crisis-prone Crisis-prone Not crisis-prone Crisis-prone Not crisis-prone Not crisis-prone Crisis-prone Not crisis-prone Crisis-prone Crisis-prone Crisis-proneCrisis entries: Crisis entries: Crisis entries: Crisis entries: Crisis entries: Crisis entries: Crisis entries:El Salvador 1981 Venezuela 1983 Guatemala 1986 Pakistan 1998 Algeria 1991 Chile 1993 Jamaica 1981, 1987Uruguay 1983 South Africa 1976 Paraguay 1986 Mexico 1982, 1995 Tunisia 1991 Egypt 1984Russia 1998 Argentina 1982 Panama 1983 Bolivia 1986Argentina 1995 1989, 1993 Thailand 1997 Peru 1978Thailand 1981 1989, 1993 Indonesia 1997 Ecuador 1982,1999Peru 1976, 1983 Philippines 1983 Uruguay 1987Brazil 1998, 2001 Argentina 2001 Indonesia 2002Jamaica 1978 Jordan 1989 Bolivia 1980Turkey 1978 Morocco 1986 Morocco 1983Korea 1997Trinidad &Tobago1988Costa Rica 1981 Turkey 2000South Africa 1985Uruguay 1990Ukraine 1998 Brazil 1983Dominican Rep 1981 Venezuela 1990,1995Sources: IMF; Standard & Poor’s; World Bank; and authors’ calculations.


- 43 -Figure 2: Rolling Tree Variable Importance12010080604020External <strong>Debt</strong>Short Term <strong>Debt</strong>PubExt<strong>Debt</strong>/RevYearsNextPresElRealGrowthExchOvervInflationExtFin'lReq'tsUSTBillsExchR_volat019891990199119921993199419951996199719981999200020012002Notes: Total External debt/GDP is replaced by Total External <strong>Debt</strong>/Exports in 1995, and by Public External<strong>Debt</strong>/GDP in 1996,97. Short Term <strong>Debt</strong> is Short Term <strong>Debt</strong>/ GDP in 89, 90,91,92,93,94,95,96,97,01 and Short Tern<strong>Debt</strong>/Reserves in 98,99,00,02. Public External <strong>Debt</strong>/GDP is replaced by Revenue/GDP in 1992,93, 98,99Figure 3: Rolling Trees’Thresholds


- 44 -0,6002,5000,5000,4000,3000,2000,1000,00019891990199119921993199419951996up to year1997199819992000200120022,0001,5001,0000,5000,000TotExt<strong>Debt</strong>/GDPInflationRealGrowthPubExt<strong>Debt</strong>/Rev(right scale)ExtFin'Ireq's (right scale)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!