Sovereign Defaults: The Price of Haircuts

That theory predicts default should be costly in credit markets and thirty years of empirical work has failed to find the cost. “Leading theories in international finance assume or predict that sovereign defaults result in higher subsequent borrowing costs, up to the government’s full exclusion from capital markets. However, empirical support for this proposition is weak at best, as shown by 30 years of research. According to the consensus of empirical studies, defaulting countries do not face substantially higher borrowing costs after a debt crisis, and often regain access to credit after just one or two years.” The authors quote Panizza, Sturzenegger and Zettelmeyer’s summary of the state of play – “None of the default punishments that the classic theory of sovereign debt has focused on appears to enjoy much empirical backing. Capital market exclusion periods are brief; effects on the cost of borrowing are temporary and small” – and Lindert and Morton’s early verdict that “investors seem to pay little attention to the past repayment record of borrowing governments.” The title of the paper is a direct answer to Bulow and Rogoff’s line that “debts which are forgiven will be forgotten.”

Measurement: a binary default indicator discards the variation that matters. “Previous papers attempting to gauge the effects of defaults on subsequent market access have used a binary default indicator, capturing any missed payment as explanatory variable for past credit history. But using binary default instead of actual losses ignores the large variation in restructuring outcomes. This may be one reason why past research concluded that the effects of previous defaults in sovereign credit markets are negligible, at least in the medium run. Our analysis indicates that it is crucial to consider the magnitude of past defaults, not only the default event per se.” The authors are explicit that they keep the econometric machinery of the earlier literature and change only the regressor: “we build on the econometric models used in 30 years of previous work on the issue, which tends to reject the claim that sovereign defaults have lasting, substantial effects on credit markets. Here, we reassess this consensus finding, with similar methods but more refined data.”

All 180 completed distressed restructurings of medium- and long-term external sovereign debt with foreign private creditors, 1978-2010, in 68 countries. Five criteria: sovereign, meaning public or publicly guaranteed debt, so privately negotiated workouts coordinated by a government (Korea 1997, Indonesia 1998) are out; distressed, following Standard and Poor’s, so “market operations that are part of routine liability management, such as voluntary debt swaps” are out; with foreign private creditors – London Club banks and foreign bondholders – so predominantly domestic restructurings and all official/Paris Club deals are out, with two exceptions (Russia’s July 1998 GKO and Ukraine’s August 1998 OVDP exchanges) admitted because they mainly hit external creditors; medium- and long-term debt, excluding 90-day rollovers and sub-year maturity extensions; and actually finalised, dropping agreements that were never implemented. The count arithmetic is worth noting: 182 restructurings are identified since 1978 – with “no restructurings occurred between 1970 and 1977 – a nonoccurrence that we find informative” – of which Togo 1980 and 1983 lack sufficient data, leaving 180. Sources: “all 29 publicly available lists on restructuring terms and more than 160 further sources, including the IMF archives, books, policy reports, offering memoranda, private sector research, and articles in the financial press.”

Because the face-value measure ignores maturity extension entirely and the market measure assumes all old debt falls due at once. The face-value haircut “captures the share of debt forgiven in nominal terms” but “ignores any loss due to a lengthening of maturities, so that a debt rescheduling, which only postpones payments in the future, implies a zero haircut” – which would mean zero loss in all 123 pure reschedulings since the 1970s, “including most cases of the 1980s and recent debt exchanges in Ukraine 2000 or Uruguay 2003. It is unrealistic to conclude that investor losses were zero in all of these cases.” The market haircut compares the present value of the new debt to the face value of the old, whose implicit counterfactual the authors illustrate with Greece’s March 2012 exchange of about 200 billion euro of bonds maturing out to 2054: taking the whole stock as due on default means the measure “will tend to exaggerate the actual loss of investors participating in the exchange.” They also dispose of the standard defence – that acceleration clauses make old debt immediately due – by pointing out that “of the 17 recent debt exchanges in Table 2, 7 were preemptive, meaning that the restructuring was implemented prior to a formal default that could have triggered the acceleration of payments.” The SZ haircut discounts both legs at the same exit yield, “which reflects the increased debt servicing capacity resulting from the exchange itself,” and has the further advantage that it “better captures the cumulative investor losses in a sequence of restructurings by the same country,” empirically relevant “as many debtor countries restructured the same debt two or three times during the 1980s and early 1990s.”

That it prices the loss at exit from default, which need not equal the loss against pre-crisis bond values, and can err in either direction. “It is important to underline that H_SZ captures the wealth loss of participating investors at one point in time, namely at the exit from default.” The alternative – measuring against the value of the old bonds before default – is rejected on a clean logical ground: “if the market is informationally efficient, the pre-default market value of the old bond will equal the expected present value of the new bond resulting from the exchange, so that this haircut would be the market’s pricing error (i.e., zero on average), which is unsatisfactory.” The authors accept the residual imprecision explicitly: the post-restructuring prices “are a natural benchmark, acknowledging upfront that this measure may overestimate or underestimate the investor’s wealth loss relative to old bond values in particular cases.” They illustrate both directions: Russia’s largely unexpected August 1998 default had a June 1998 pre-default price above the value they compute for the old bonds at exit, while Uruguay 2003’s pre-crisis prices were below.

By imputing a deal-specific rate from US low-grade corporate yields and the sovereign’s rating, because observed exit yields exist only for a few recent cases and a constant rate is indefensible. Prior work either used market exit yields, available “only for a very small subsample of recent cases with liquid secondary debt markets,” or a constant rate – commonly a flat 10% – “despite the fact that countries restructured their debts in very different conditions.” The paper’s illustration of why that fails is concrete: Nigeria’s credit rating at its 1991 restructuring was 19.5 on the institutional investor scale against South Africa’s 38.2 in 1993, and the speculative-grade US corporate yield was 11.43% when Russia restructured in August 2000 against 8.14% when Argentina restructured in 2005 – “our procedure takes into account both of these factors and gives different yields for these four cases: 9.81 for South Africa, 10.36 for Argentina, 12.48 for Russia, and 18.28 for Nigeria.” The procedure yields monthly rates “from the late 1970s until 2010 and for up to 140 countries”; in the subset actually used they “range between 9 and 41.2 percent, with a median of 15,” with quartiles at 12.8 and 24.3. Cash flows are computed in US dollars loan-by-loan and bond-by-bond wherever the terms could be collected, and in aggregate where they could not, “as often happens in restructurings of the 1970s and 1980s.”

Average 37%, enormous dispersion, rising over time, and far below the corporate benchmark. The mean SZ haircut for 1978-2010 is 37.04% (SD 27.28, range -9.8% to 97%); the volume-weighted average is about 30%, implying “that, on average, investors could preserve almost two-thirds of their asset value in restructurings of the past decades.” The market haircut averages 40.01% and face-value reduction only 16.77%. For cases where the old debt had not fully come due, “the average H_SZ is 6.5 percentage points lower than the average H_M,” with the gap reaching 22 percentage points. Dispersion: “one-half of the haircuts are either below 23 percent or above 53 percent,” and it has widened – “recent years have seen a particularly large variation, with some deals involving haircuts as high as 90 percent and others involving haircuts as low as 5 percent.” By era the means are 25.57% (1978-1989, 99 cases), 51.81% (1990-1997, 48 cases) and 49.96% (1998-2010, 33 cases): “average haircuts were about 25 percentage points higher during the 1990s and 2000s as compared to deals implemented during the 1970s and 1980s,” which the authors attribute to 1980s deals mostly extending maturities and “thus postponing the day of reckoning that many debtor countries had deep-rooted solvency problems.” The 17 Brady deals average 45%, exceeding the 39% mean for the 17 bond restructurings since 1998. Debt reduction is what moves the number: the 57 deals with face-value write-offs average 65% against 24% for the 123 pure reschedulings. Debtor type matters most of all: the 23 donor-supported HIPC restructurings co-financed by the World Bank’s Debt Reduction Facility average 87.03%, “nearly three times as large as for restructurings in middle income countries” (29.72%). The sovereign average of 37% sits against Moody’s estimate of 64% for US corporate bond and loan restructurings 1982-2005, “nearly twice as high as what we find for sovereign debt,” a discrepancy the authors call surprising given that corporate debt is court-enforceable and subject to an orderly bankruptcy regime. Negative haircuts appear in a small set of mostly early-1980s cases, arising “from a restructuring in which the interest rate on the new debt exceeds the estimated discount rate prevailing at the time,” so lengthening maturities raises rather than lowers present value.

Very close to Sturzenegger and Zettelmeyer and the central-bank estimates; far from Benjamin and Wright. For the overlapping recent cases the mean absolute deviation from SZ’s average haircuts is 5.8 percentage points, with only Pakistan 1999 and Ukraine 2000 differing by more than 10 points, “mostly because our methodology yields significantly lower discount rates for these two cases.” Against the Bank of England, Bank of Spain, IMF and rating-agency estimates, “for each of these set of estimates, the correlation with our preferred measure exceeds 90 percent.” Benjamin and Wright are the outlier: correlation 0.54 and mean absolute deviation 21 percentage points against the SZ haircut, and correlation 0.49 with MAD 26 points against the paper’s own face-value measure. The authors give three reasons and conclude the two are “not directly comparable”: Benjamin and Wright use World Bank GDF debt-relief flows rather than event-based restructuring terms, GDF “do not differentiate between debt relief by private creditors and by official creditors,” and GDF data are debtor-reported, which the authors flag via Depetris Chauvin and Kraay’s caution that “given the weak debt management capacity of many low-income countries and the complexity of many debt restructurings, we expect debtor reported data on debt relief to be relatively noisy.”

No, and the paper says so. The two tested predictions are “the larger the size of H, the higher the yield spreads after restructurings; and the larger H, the longer the period of exclusion from capital markets,” but “these predictions are not derived from an established theory, among other things because most models simply assume a world with zero recovery rates. Hence, we do not explicitly test a specific theoretical model.” The authors situate the results against several candidate channels without adjudicating: the Eaton-Gersovitz exclusion punishment; a high haircut as a signal of “untrustworthy economic policies, expropriative practices by the incumbent government” (Cole-Kehoe-style) or of “negative private information held by the government about purely economic fundamentals” (Sandleris-style); and the Grossman-van Huyck distinction under which adverse effects should arise only from an “inexcusable” haircut not justified by bad exogenous conditions – an extension the authors report having pursued in an earlier version and omitted here for brevity.

Two, in opposite directions: that high-haircut countries are simply in worse shape, and that debt relief itself should make them safer. On the first: “it is possible that countries imposing higher haircuts are also in worse shape than those imposing lower haircuts. Country characteristics could influence both the size of H and country access conditions after the restructuring.” The response is country and time fixed effects plus “a large set of observable, time-varying fundamentals suggested by theory and the previous international finance and asset pricing literatures” – debt/GDP, reserves/imports, inflation, real growth, current account, primary balance, a low-grade US corporate yield index for global conditions, a residual from regressing credit ratings on fundamentals (following Eichengreen-Mody and Dell’Ariccia et al. rather than including ratings directly), the ICRG political risk index, years in office, and a new-government dummy. The authors’ verdict on this is deliberately unsatisfying: “This mitigates, but not necessarily completely eliminates, the possibility that the coefficients on H may pick up the effect of a confounding variable which remains omitted.” On the second: “Sovereigns imposing high haircuts will reduce their indebtedness more significantly, lowering the debt to GDP ratio and possibly decreasing the likelihood of future default, at least in the short run. Thus, in an atomistic bond market without creditor collusion lenders could ultimately reward sovereigns for imposing high haircuts,” which would push the coefficient the other way; they control for post-restructuring debt/GDP and the sovereign rating.

With a binary dummy, the received short-lived result; with the continuous haircut, an effect that persists to year seven. The panel is monthly EMBI Global stripped spreads over US Treasuries for 47 countries, January 1993 to December 2010, over 5,000 observations – 23 defaulters and 24 non-defaulters – with observations during default excluded and standard errors clustered on country-year. Only 27 of the 67 final restructurings, from 23 countries, fall inside EMBIG coverage. The dummy specification gives 262.54 basis points in year one and 151.23 in year two, then coefficients that are “small and/or only marginally significant thereafter,” confirming that “with a binary measure of default, we confirm the results of the received literature that default effects appear very short-lived.” Substituting the haircut (column 3, country and year fixed effects) gives 6.75 basis points per percentage point of haircut in year one, 4.73 in year two, 3.89 in year three, 3.16 in years four and five, and an insignificant 0.80 in years six and seven – so a 40% haircut, roughly the EMBIG-sample mean, maps to “270 bp higher spreads in year 1 and 127 basis points higher in years 4 and 5,” and a one-standard-deviation (22 percentage point) increase to 149, 104, 85 and 70 basis points in years one, two, three, and four-to-five. With the full battery of controls (column 4) the coefficients are “somewhat less pronounced” but “remain economically and statistically significant up to year seven.”

Because the haircut-only specification cannot separate the cost of defaulting from the cost of defaulting deeply, and because it omits the constitutive term of an interaction. “The advantage of estimating equation (4) is that it allows us to disentangle the spread increase associated with the default per se from the spread increase associated with the size of haircuts (occurrence versus magnitude).” Methodologically the haircut-only model “is methodologically problematic because it only includes the interacted variables (the lagged H_i s) but not the constitutive terms (the lagged r_i s),” which Brambor, Clark and Golder say should always be included. F-tests confirm both groups belong: 5.46 for the lagged haircuts and 4.54 for the lagged dummies in column 5, both with p-values below 1%. The authors then warn that the individual coefficients in the full model “cannot be taken at face value, but have to be interpreted conditionally,” so columns 5-8 are not comparable to columns 2 and 4, and that the coefficient changes from adding r “should not be seen as the result of multicollinearity.” On multicollinearity itself they are explicit rather than defensive: the correlation between r and H is about 0.85, which inflates individual standard errors but “actually lowers the variance of the estimated effect of interest,” the sum of the two coefficients – “the high correlation between r and H complicates making inference about their individual effects, but facilitates inference about their sum.”

Statistically significant for haircuts above 40% in every year from one to seven, with a one-standard-deviation increase worth 122 basis points in years four and five and 149 in years six and seven. The interpretation runs through the conditional spread increase, the sum of the haircut coefficient times H and the dummy coefficient, plotted with 95% confidence bands for each lag: “restructurings are statistically significant for years 1-7, and for haircuts above 40 percent… (lag 3 is the only exception, with significance reached above H = 55 percent).” The authors set the magnitude against the prior literature deliberately: “the influential early studies by Lindert and Morton (1989) and Özler (1993), and newer papers like Benczur and Ilut (2009) or Catão, Fostel, and Kapur (2009) suggest that a past default leads to an average increase in post-crisis spreads of, at most, 50 basis points. So while defaults may seem costless when not controlling for investor losses, we find that large haircuts can be associated with substantially higher spreads for up to seven years afterward.” The effect is strongest late: “The spread increase associated with haircut size is also economically substantial, especially for years four to seven after a restructuring.” Robustness is reported as holding under alternative specifications and samples, with yearly rather than monthly data, with alternative haircut measures including cumulative haircuts, and when dropping the three countries with repeated final deals.

As years from a final restructuring to the first year of either an international placement that raises public indebtedness or positive net public-sector transfers from private foreign creditors. The measure deliberately merges the two strands of prior work – individual loan and bond issuance (Eichengreen-Portes; Gelos-Sahay-Sandleris) and aggregate capital flows (Richmond-Dias) – “to avoid lengthy discussions on the benefits and drawbacks of alternative definitions and data sources.” Issuance data come from Dealogic 1980-2010, aggregating “8,776 individual public and publicly guaranteed bonds in 95 developing countries and 10,212 public or publicly guaranteed syndicated loans from 136 countries,” restricted (following Gelos-Sahay-Sandleris) to issuances that raise public indebtedness; flow data come from GDF net bank and bond transfers, with arrears not counted as a positive transfer. Crucially, the object is exclusion after the crisis is resolved: “we study market exclusion after a default is successfully resolved, not exclusion during default.” Robustness variants include “full reaccess” (flows above 1% of GDP), an issuance-only measure, and a public-plus-private-sector measure.

A large, robust association: higher haircuts, much slower reaccess. Across the 67 final restructurings, “the average duration from restructuring to partial reaccess is 5.1 years, while the median is three years,” but split by haircut, “partial reaccess takes just 2.3 years after cases with H_SZ < 30 percent, while the duration is more than twice as long (6.1 years) for cases with H_SZ > 30 percent.” Kaplan-Meier survival functions split three ways make the tail vivid: “Countries with H_SZ < 30 have a 60 percent probability of reaccessing markets within three years compared to just 10 percent for countries with H_SZ > 60,” and “countries imposing H_SZ > 60 percent are likely to remain excluded after 10 years, with an unconditional probability exceeding 50 percent.” The multivariate estimate is a semi-parametric Cox proportional hazard model – chosen because it leaves the baseline hazard unparameterised and can handle censoring and repeat events, with the Lin-Wei variance correction for countries with multiple episodes. The haircut coefficient “is negative and robustly significant in all specifications”: -0.024 in the full model, meaning “a 1 unit (percentage point) increase in H_SZ lowers the likelihood of reaccessing capital markets in a given year by 2.4 percent,” so “according to our most conservative estimate, a 1 standard deviation increase (30 percentage points in this sample) is associated with a 51 percent lower likelihood of reaccess in any given year.” Results hold when HIPCs are excluded. Among the controls, only population, GDP per capita and a good credit rating predict quicker reaccess, with debt/GDP and the fiscal balance sometimes negative, while “political risk, annual inflation and growth, or the ratio of reserves to imports are clearly insignificant.” The authors note that Richmond and Dias, using Benjamin and Wright’s haircut data, “do not find haircuts to be a significant predictor for partial reaccess.”

The mechanism, and any prescription about minimising creditor losses. “Our results should, however, not be misinterpreted. We did not identify a direct channel linking haircuts and sovereign borrowing conditions. Thus, we cannot be sure whether we observe punishment effects, reputational effects or neither of the two. The results also do not imply that countries in default should try to minimize creditor losses. Instead, we provide indicative evidence for the existence of a trade-off; achieving a high degree of debt relief now can have benefits in the short-run, but may also imply worse borrowing conditions in the future.” Earlier, the same caution is stated in the conditional: “While higher borrowing costs and market exclusion could be the result of punishment effects or of information revelation, our findings should not be interpreted as direct evidence for any of the two.” The stated agenda for future work is correspondingly narrow: “we see the need to study the mechanisms behind our results, both empirically and theoretically,” and to “assess the determinants of high or low haircuts.”

Disciplining theory, judging burden sharing, and supplying sovereign recovery rates that did not previously exist. “On the academic front, they invite us to rethink the influential theoretical models that feature a 100 percent haircut upon default. On the policy front, they enable more informed judgments on debt crises outcomes and private creditor burden sharing in the past decades.” And on asset pricing: the dataset “provides, for the first time, representative estimates on sovereign debt recovery rates,” usable “as inputs for a wide range of credit risk models in the financial industry, e.g., to back out default probabilities from observable bond prices.” The gap being filled is real: the authors note that “given the lack of data, even rating agencies continue to base their recovery assumptions for sovereigns on a very small sample of restructurings” – 15 cases in the most recent Moody’s report, 10 in Standard and Poor’s.