Sovereign Bond Contracts: Flaws in the Public Data?

Abstract

Commercial databases now make available to paying clients information about the legal terms in sovereign loan contracts. This information is important to academic researchers, to policy institutions such as the International Monetary Fund, and to investors and other market actors. For a random sample of 10 countries, the authors compare this data to a hand-coded sample of bond terms. They find significant error rates in the commercial databases, which vary significantly by country and by the legal term at issue. In some cases, they document error rates well over 75 per cent. They also describe important limitations in the data, especially the use of binary coding schemes that obscure important differences in the rights conferred by different sovereign loan contracts.

I. INTRODUCTION

For years, the economic literature on sovereign debt paid little attention to the legal terms in loan contracts.¹ This indifference derived from the law of sovereign immunity, which effectively barred national courts from hearing lawsuits and enforcing judgments against foreign sovereigns. In the 1970s, however, the law of sovereign immunity changed, so that sovereigns engaged in commercial activity on foreign soil became subject to some degree of legal coercion.² No immediate wave of litigation followed, but creditor lawsuits have gradually become commonplace features of the sovereign debt landscape.³ Today, almost every restructuring will prompt litigation, often by creditors holding claims worth many billions of dollars.

Now, the legal terms in sovereign loan contracts matter. After Argentina’s restructuring in the mid-2000s, holdout creditors invested tens of millions of dollars in litigation, winning important legal victories that turned on a close reading of the relevant bond contracts. For some creditors, this strategy produced returns on investment estimated at over 800 per cent.⁴ Demand for information about sovereign loan contracts—which are typically bond contracts—has generated supply. Commercial databases now summarize key contract terms for paying customers and, in some cases, include the documentation underlying the loan. These databases potentially let market actors distinguish bonds on legal as well as economic grounds. The databases can also facilitate large-scale studies of loan contracts by academic researchers and public sector institutions such as the International Monetary Fund (IMF). Previously, these studies would have been difficult or impossible to complete.

We have sometimes used these databases in our own research. It is time consuming to gather loan contracts and to code them by hand—although we have spent many hours doing just that. For this reason, many studies of sovereign debt contracts involve small samples. With the advent of commercial databases, users can quickly download information about thousands of contracts in a form suitable for statistical analysis.

These potential benefits, however, depend vitally on whether the databases provide quality data on the terms of loan contracts. In this paper, we take a preliminary look at this question. For a subset of countries, and using the two most prominent commercial databases, we ask two questions. First, how accurately do these databases report contract terms? Second, is what they report useful—that is, do they capture legally relevant differences among contracts?

We find that the quality of data leaves something to be desired. To summarize:

• Commercial databases have significant error rates. Averaged across countries, some legal terms are mis-coded nearly 50 per cent of the time. Within countries, error rates occasionally exceed 75 per cent.

• The databases typically report only binary information about whether the loan contract includes some version of a clause. But clauses differ in legally relevant ways; binary coding obscures these differences.

• The databases report some details that are irrelevant, while omitting much relevant information.

This article proceeds as follows. Section II briefly examines the degree to which academics, market actors, and the primary policy institution in this area (the IMF) use the data provided by the commercial databases. For a randomly selected sample of 10 countries, section III focuses on how the two primary databases code a number of important legal terms, comparing this to a hand-coded dataset based on the relevant sales documents.⁵ Section IV asks whether commercial databases report information that, even if accurate, is useful. Section V concludes by discussing the implications of these findings.

II. COMMERCIAL DATABASES ARE IMPORTANT TO ACADEMICS, POLICYMAKERS, AND MARKET ACTORS

1. Academic papers

The empirical study of contract terms is a relatively new field, in part, we suspect, because good data is hard to find. For sovereign bonds, the difficulty is compounded by the fact that most loans consist of multiple documents, most of which are not readily available even to the investors to whom these bonds are marketed. In recent years, thanks to firms like DCM and Bloomberg, the number of studies has increased. To get a rough sense of the importance of DCM and Bloomberg to recent empirical work on sovereign debt contracts, we identified articles published in the last five years (1 January 2013 through to 1 May 2018) that meet two criteria: (i) the article reports information about legal, not merely financial, aspects of sovereign bond contracts; (ii) some or all of this information could have been gathered from DCM and/or Bloomberg (whether or not the authors in fact used these sources).

We found 20 published articles.⁶ In all but six, the authors used data from either Bloomberg or DCM (and sometimes both), either as a primary source or to supplement data gathered from other sources. Many of these articles, however, use Bloomberg and DCM primarily for financial data, rather than for the coding of legal variables. When academic researchers do use commercial databases for legal terms, they tend to focus on two clauses. The first clause specifies the law that will govern the bonds. The second permits a bondholder majority or super-majority to amend key bond terms and to impose this result on dissenting bondholders. This latter clause is often called a collective action clause, or CAC. As we report in the next section, the commercial databases are relatively accurate in their coding of governing law clauses. However, there are often significant inaccuracies in the coding of CACs.

It appears that academic researchers make only limited use of the legal terms reported by commercial databases. As an initial matter, this is puzzling, as the databases make it easy to construct large-N samples of bond contracts that include information about both legal and financial terms. Of course, the data is relatively new, so one might expect a delay before researchers discover and exploit it. But it is also possible that some researchers have concerns about data quality. As we show in section III, these concerns may be warranted.

2. Informing official sector policy

For nearly a century, governments and other policy actors have worked to influence the legal terms in sovereign loan contracts.⁷ Many of these efforts have been motivated by concerns about the collective action problems inherent in bond lending. For example, individual bondholders may refuse to participate in a restructuring that would benefit the group as a whole, hoping to gain a better deal for themselves. That prospect may complicate and delay a restructuring. Once a restructuring occurs, holdouts may also enforce their claims in ways that harm participating bondholders, such as by trying to seize payments made under the restructuring plan. Official sector actors have repeatedly sponsored initiatives designed to encourage the use of contract terms (such as CACs) that mitigate these collective action problems. In the modern era, the IMF is perhaps the key player in this area.

We looked for reports and working papers issued by the IMF and the World Bank (another Bretton Woods institution) over the past five years and which discuss the terms of sovereign bond contracts. We found five reports, all by the IMF.⁸ These generally assess current market practice and trends, focusing on contract terms that the IMF views as important to its policy goals. In particular, the IMF has recently promoted the use of ‘aggregated’ CACs. Traditional CACs call for a restructuring vote to occur within each series of bonds issued by the government. Because governments have many such issuances, and some are relatively small, it is easy for prospective holdouts to buy a position large enough to block a restructuring of that series. Aggregated CACs reduce this risk by aggregating the vote across multiple series of bonds.

Each of the IMF reports uses data extracted from commercial databases, including Bloomberg and DCM. Occasionally, however, the IMF appears to recognize the limits in the coding made available by these databases. Most notably, the IMF contracted specifically with the Perfect Information database to develop a custom coding scheme to gather information about the use of aggregated CACs. This decision was apparently motivated by the fact that existing databases did not distinguish between traditional CACs and CACs with aggregation features. (Indeed, DCM does not appear to code for CACs at all.) We will have more to say about such coding ‘gaps’ in section IV. For now, we note only that, while official sector institutions may recognize gaps in commercial databases, we have seen no indication that they doubt the accuracy of the data being reported.⁹

3. Reliance by market actors

Next, to understand better how market analysts use the coding from Bloomberg and DCM, we collected research reports addressing the ongoing sovereign debt crisis in Venezuela. The contract terms in Venezuela’s loan contracts vary in a number of ways, and these differences have generated much discussion. Because these reports are not publicly available, we drew on industry contacts to gather as many as we could. We collected nine reports devoting significant attention to the terms of loan contracts.¹⁰ Each depends almost exclusively on Bloomberg for financial data. Interestingly, these market reports—like the IMF—appear to rely less on Bloomberg for the coding of legal variables.

As an example of the discussion generated by differences in the legal terms of Venezuela’s bonds, many analyst reports emphasize the importance of CACs (which, recall, specify the voting threshold for modifying loan terms). A sovereign bond typically specifies one voting threshold for proposals to modify payment-related terms, such as the amount of interest and principal payments, and another threshold for proposals to modify terms unrelated to payment. For Venezuela, however, the same proposal will trigger different voting thresholds, depending on which bond an investor holds. For example, for proposals to modify payment-related terms, some bonds allow a 75 per cent super-majority to accept on behalf of all holders of the affected series.¹¹ Others require an 85 per cent super-majority. Still others require unanimity; holders of these bonds need not worry that a bondholder super-majority will modify payment terms against their wishes. Thus, investors in Venezuelan debt can buy some protection from restructuring by acquiring a bond with a higher voting threshold. Market analysts took pains to emphasize this variance in their research reports.

For example, a Morgan Stanley research report in January 2018 included nine separate tables reporting pricing, payment, and other financial information gathered from Bloomberg. Perhaps tellingly, when compiling a table describing the bonds’ legal terms (the presence of CACs and the CAC thresholds), the report’s authors did not use Bloomberg. Instead, they hand-coded the CAC thresholds from the bond documents. Similarly, JPMorgan’s ‘Venezuela and PDVSA Debt: A Guide’ (March 2015) used data from Bloomberg for financial information about the bonds, yet hand-coded the bond prospectuses for two tables referencing the governing law, CACs, and cross-default provisions.

Deutsche Bank appeared to trust Bloomberg’s legal coding more. One report (‘Venezuela: Preparing for the End Game’) includes a chart with both financial and legal details of bonds issued by Venezuela and state-owned oil company PDVSA. The chart uses data both from Bloomberg and from hand-coding of bond prospectuses. It appears Deutsche Bank used Bloomberg’s legal coding for the CACs, as the data provided in the table tracks that reported by Bloomberg (including its errors). However, Deutsche Bank clearly hand-coded some information, as the report discusses some legal terms, such as the ‘grace period’ in which the sovereign may cure a default, that neither Bloomberg nor DCM covered. We will return to this and similar omissions from commercial databases in section IV.

III. ERROR RATES IN THE COMMERCIAL DATABASES

Thus far, we have seen that academics, policymakers, and market actors studying sovereign debt contracts rely selectively on the commercial databases. We see widespread reliance for financial information. We see much less reliance for legal information, although there are occasional exceptions, which typically involve reliance on commercial databases for information about governing law and the presence of a CAC. In this section, we ask whether consumers should be relying on commercial databases for information about legal terms.

To investigate, we first downloaded loan covenant data from Bloomberg and DCM for 10 randomly selected countries.¹² The Bloomberg sample consists of 411 bonds that had not yet reached maturity.¹³ Bloomberg designated 284 of these as international bonds issued on foreign capital markets. These are typically governed by foreign law, enforceable in foreign tribunals, and repayable only in foreign currency. The rest were issued on domestic capital markets. Most such bonds are subject to the issuing government’s law and are repayable in its currency. Bonds of this sort are notably informal; they rarely include the extensive contractual protections found in international bonds. Nevertheless, we include them to check the accuracy of the coding of the governing law. The DCM sample consists of 1350 bonds, of which roughly 580 were international issuances.¹⁴ The DCM sample is larger, primarily because it includes matured as well as unmatured bonds.¹⁵

To develop a comparison sample, we began with documents gathered in the course of constructing a larger dataset spanning nearly two centuries. For the modern era of bond lending, that dataset includes 1600 bond issuances by 147 sovereign and sub-sovereign entities, including 227 issuances from the 10 governments we discuss here.¹⁶ However, our data overlapped imperfectly with that reported by Bloomberg and DCM.¹⁷ We thus consulted commercial databases, such as Perfect Information and Thomson One Banker, which sometimes include sales documents or (rarely) loan contracts. We gathered as many sales documents as we could find from these sources, resulting in documentation for a total of 392 issuances.

After reading the terms described or reprinted in the sales documents, we hand-coded information about a number of legal terms also reported by Bloomberg and DCM:

• Governing law: specifies which jurisdiction’s law will govern disputes over the meaning and enforcement of the bond;

• Cross default: specifies when a default on other obligations will also constitute a default on the relevant bond;

• Force majeure: specifies whether and when the government’s obligations may be excused or deferred due to circumstances outside its control, such as war or natural disaster;

• Negative pledge: forbids the government to subordinate bondholders formally by creating new, senior classes of debt;

• Pari passu: forbids certain kinds of informal subordination—for example, in some circumstances, selective payment of some bondholders but not others might violate the pari passu covenant;

• CACs: specify voting threshold for modifying loan terms.

We then compared this hand-coded dataset with the coding provided by the commercial databases. Because of omissions in the data reported by Bloomberg and DCM, our comparison samples vary in size depending on the clause, as shown in Table 1.

1. Discrepancies in the Dealogic coding

Dealogic’s coding overlaps with our dataset on the governing law, cross default, force majeure, and negative pledge clauses. Dealogic’s overall error rate, across all of these clauses, was 32.4 per cent. It most accurately coded governing law, with an average error rate—ie averaged across countries—of 4.6 per cent. In contrast, there were frequent inaccuracies in its coding of the cross default and negative pledge clauses, with respective average error rates of 45.3 and 49.7 per cent. The error rates also varied across countries, as shown in Figure 1.

There is no readily discernible explanation for this pattern of errors. When coding for the presence of a negative pledge clause in bonds issued by Finland, for example, Dealogic’s error rate was only 4.9 per cent. The error rate was much higher for other countries; for seven of the ten, the error rate was over 50 per cent. Likewise, when coding addressed the presence of a cross-default clause, the error rate exceeded 75 per cent for five of the ten countries.

2. Discrepancies in the Bloomberg coding

The Bloomberg database covers a wider range of contract terms, including terms that we view as important to market participants, academics, and policy actors. Thus, there are more opportunities to compare our dataset to Bloomberg’s. On average, Bloomberg’s coding was also more accurate, with an overall error rate (across all countries and legal terms) of 8.9 per cent. But we still find relatively frequent inaccuracies in Bloomberg’s coding of some clauses. For example, we find an error rate of 28.7 per cent for the presence of a pari passu clause, and an error rate of 10.1 per cent for the presence of a CAC. As with DCM, Bloomberg’s error rates varied by country in ways that we cannot readily explain. Figure 2 shows this, focusing on pari passu, collective action, and cross-default clauses, for which Bloomberg had relatively high average error rates.

For comparison purposes, Figure 3 reports average error rates, by clause, for the two commercial databases. Note that DCM does not code for the pari passu clause or for CACs.

3. Do the errors matter?

Are these errors common enough, and distributed non-randomly, so as potentially to bias studies relying on data from Bloomberg and DCM? We are not yet in a position to offer a definitive answer, but there is reason for caution. Neither Bloomberg nor DCM explains its coding method in enough detail for us to identify the source of the errors. That said, the error rates depicted in Figures 1–3 are too high, and too varied, to be easily discounted. In our own work, relying on hand-coding, we typically find error rates in the range of 1 to 5 per cent. But the error rates in the commercial databases are often much higher, occasionally near 50 per cent. The fact that error rates vary dramatically by country and by clause also implies that errors are not randomly distributed. The end result, we think, is that the data on these legal variables—except perhaps for governing law—are essentially unusable in academic or policy research.¹⁸

Errors of this sort have potentially important consequences. Academic research can be used to motivate and justify policy action. For example, official actors have repeatedly encouraged wider use of CACs to facilitate debt restructurings. But many governments hesitate to adopt these clauses, fearful of suffering a pricing penalty if investors interpret adoption as a signal of impending default. To inform this discussion, a number of academic papers have addressed the pricing consequences of CACs; flaws in the underlying data can skew the debate.

Given the preliminary nature of our inquiry, we are not yet in a position to ask whether corrected data might change the results reported in the literature. But it is certainly possible. As an example, we use an article co-authored by one of us as a foil. The goal in the article was to examine the impact of contract terms on sovereign bond pricing in a crisis.¹⁹ The focus was Venezuela, which had spent years in deep crisis and also had a large debt stock with considerable variation in the underlying contract terms (a rare, if not unique, combination).

The study addressed the pricing effects of two clauses, CACs and pari passu. For background, although the IMF and other official actors have long encouraged the use of CACs, they have recently encouraged sovereigns to revise these clauses to reduce the likelihood of holdout creditors and to eliminate certain disruptive litigation tactics. Once again, studies examining the pricing effects of such a change have proven important. In advocating for the revisions, the IMF cited empirical studies that found little evidence that CACs have negative pricing effects.²⁰ Yet these were multi-country studies, where almost no country was in a debt crisis. And it is in crisis, when the threat of litigation is real, that investors are most likely to assign value to contract terms affecting litigation rights.

The paper we discuss here began with an analysis based on Bloomberg’s coding of CACs and pari passu. The initial inquiry supported the official story that the market did not impose a price penalty on bonds that were easier to restructure (and harder to litigate). However, this preliminary result conflicted with reports from market contacts, who asserted that activist hedge funds were targeting bonds that were harder to restructure. Further investigation revealed errors in the Bloomberg coding that mostly affected the pari passu variable. The authors corrected these errors, and also replaced Bloomberg’s binary coding scheme with one that better captured legal differences among the bonds. The result: bonds that combined the hardest-to-restructure CAC with the most litigation-friendly pari passu clauses were priced at a premium. In other words, the results conflicted with prior research and arguably undermined the basis for the IMF’s policy reform.²¹

IV. DO COMMERCIAL DATABASES PROVIDE USEFUL INFORMATION?

We also wanted to know whether the commercial databases provide information that is useful to policymakers, academics, and market actors. In this section, we evaluate the practice of binary coding, common to both Bloomberg and DCM, which treats contract clauses as either present or absent without accounting for textual variations that potentially impact on the rights conferred.

As a stylized example, consider the negative pledge clause, a covenant that forbids the government (subject to any exceptions specified by the clause) to subordinate bondholders by issuing new, senior debt. Now consider three bonds. Bond A has no negative pledge clause at all; these bondholders may be subordinated at the government’s whim. Bond B has a negative pledge clause with no exceptions; these bondholders may not be subordinated by the issuance of senior debt, full stop. Bond C has a negative pledge clause, with an exception that lets the government create senior debt ‘if necessary to maintain access to financial markets on reasonably acceptable terms’. This (hypothetical) exception negates much, though probably not all, of the protection offered by the negative pledge clause. Yet a binary coding scheme would ignore the crucial difference between these bonds and would likely treat bonds B and C as identical.

Of course, any attempt to create a more nuanced coding scheme encounters an immediate difficulty. If different loan contracts include different versions of a clause, then which differences matter? The question is hard to answer ex ante. When a bond is issued—ie at the time the bond is coded by commercial database providers—it may not be clear precisely which differences will prove important in a resulting debt crisis. Despite this difficulty, it is possible to form rough judgements about which differences have a relatively high probability of affecting investors’ rights.

1. Governing law clauses

Historical experience teaches that the choice of governing law matters a great deal. For instance, at the time of its 2012 restructuring, approximately 90 per cent of Greece’s debt was governed by that country’s own law. Because of this, the government easily and efficiently reduced its debt by modifying its own law to facilitate restructuring. By contrast, it had a much harder time restructuring the subset of its debt that was governed by foreign (typically English) law, which the Greek government had no power to modify.²²

Bloomberg and DCM quite sensibly treat governing law as a categorical variable, rather than as a dichotomous one. With this data, researchers and investors can quickly identify the jurisdiction whose legal rules will govern the issuer’s obligations. Unfortunately, when it comes to variables other than governing law, the commercial databases provide less useful information.

2. Cross-default clauses

Cross-default clauses specify when a government’s default on other obligations constitutes an event of default under the relevant bonds. In effect, the clause defines a set of early warning signals, which trigger an investor’s right to declare a default even if the government remains current on its obligations to them. This is valuable protection, for it allows an investor to accelerate the bond—ie to demand full payment of principal and accrued interest—at early signs of trouble, rather than wait until the government officially runs out of money to pay. It is obvious, then, that there is an important difference between a bond that specifies only one early warning signal and a bond that specifies many.

Both Bloomberg and DCM appear to treat the cross-default clause as a dichotomous variable, reporting only whether the bond includes some version of the clause. This coding scheme is unnecessarily limited. For example, many government bonds provide that it will be an event of default if the government fails to make a payment due on other foreign-currency-denominated debt, or fails to pay court judgments, so long as the unpaid amounts exceed a specified threshold. Different bonds specify different thresholds (say, $20 million or $100 million), with obvious implications for the likelihood that the cross-default clause will be triggered.²³ A useful coding scheme would capture the monetary threshold that must be exceeded before a default on other obligations triggers a bond investor’s right to declare a default.²⁴

As a second example, consider the unique relationship between governments and some state-owned entities, especially firms whose exports generate a significant share of the economy’s foreign currency reserves. State-owned oil and natural gas companies are an example. Although technically separate entities with separate debts, the state and the entity may be financially interdependent. One might therefore expect investors in sovereign debt to wonder whether default by a major state-owned enterprise will trigger the cross-default provisions in the sovereign’s own bonds (and vice versa).²⁵

3. Force majeure clauses

The commercial databases sometimes code for the presence of contract clauses that seem relatively unimportant (or at least, exceptionally rare). For example, both DCM and Bloomberg include coding for whether the bond includes a force majeure clause. As noted, this clause specifies circumstances outside the government’s control that will excuse its obligations. In theory, one can see the relevance of such a clause. Commercial contracts between private parties often include force majeure clauses, and these may excuse non-performance when caused by an act of war. One can imagine an unprovoked act of war that impairs a government’s ability to pay its debts; in such a case, a force majeure clause would allow it to suspend payments without this constituting a default.

Still, we question the necessity of coding for the force majeure clause. Based on our experience reviewing thousands of bond contracts, sovereign bonds rarely if ever include this clause. A few sovereign bonds issued several decades ago—so called ‘Brady Bonds’—linking payment obligations to exports of natural resources were subject to conditions roughly analogous to those specified by a force majeure clause.²⁶ Except for these rare exceptions, which are now old history, our hand-coded dataset includes no contracts with force majeure clauses. Indeed, we are unaware of a sovereign government ever successfully invoking force majeure to excuse its non-payment of bond debt.

4. Negative pledge clauses

There would seem to be an important difference between a loan contract with a negative pledge clause and a loan contract without one. In the former case, lenders need not worry that the borrower will subordinate them by issuing senior debt; in the latter, lenders must endure this risk. By treating the negative pledge clause as a dichotomous variable, Bloomberg and DCM do capture this difference (at least to the extent that the coding is accurate). We hope, however, that our earlier discussion will have persuaded readers that negative pledge clauses can differ in important ways, which Bloomberg and DCM make no effort to capture.

The example we used earlier—a negative pledge clause that allowed secured debt when ‘necessary to maintain access to financial markets on reasonably acceptable terms’—involved a hypothetical exception to the clause’s protection. Here is a real exception: bonds issued by Venezuelan state oil company PDVSA include a negative pledge clause that allows the company to create liens in favour of the Venezuelan government. This exception may prove important in a debt restructuring, for it apparently allows PDVSA to subordinate bondholders to the government in exchange for new financing.²⁷ Although the example is specific to Venezuela,²⁸ the lesson is a general one: The extent of a bondholder’s protection against subordination depends not on whether the loan includes a negative pledge clause, but on the number and content of the exceptions.

In theory, there might be an infinite number of such exceptions, and no coding scheme could account for them all. In our experience, however, most sovereign bonds begin with a relatively standard template. It would be feasible to investigate whether there is a relatively standard set of exceptions to the negative pledge clause. If so, coding could capture whether a bond includes these standard exceptions, and also whether the bond includes any additional exceptions (without attempting to capture the content of these additional terms). Such a coding scheme would at least alert users of the data to the presence of idiosyncratic bond terms and the need for further investigation.

5. Pari passu clauses

The meaning of the pari passu clause is disputed. At minimum, most observers understand the clause to prevent a borrower government from using informal means to subordinate one pari passu-ranking creditor to another.²⁹ More controversially, some versions of the clause have been interpreted to prohibit selective payment to a subset of pari passu-ranking creditors, at least under certain conditions.³⁰ So interpreted, the clause becomes a potent weapon for holdout creditors. For example, after Argentina’s 2005 and 2010 debt restructurings, holdout creditors persuaded US courts to issue orders preventing Argentina from servicing its restructured debt unless it also paid holdout creditors in full.³¹

Because the interpretation is contested, the precise language of the pari passu clause matters. For example, some versions of the clause provide only that the bonds will ‘rank equally’ with other unsecured debt (typically, only debt denominated in a foreign currency). Other versions, including that used by Argentina, provide that the government’s ‘payment obligations’ under the bonds will rank equally with other unsecured debt. This priority of payment language arguably expands the prohibition, so that it forbids some instances of selective payment in addition to behaviour that effectively alters the ranking of creditors. Argentina’s creditors emphasized this priority of payment language in pressing their claims, and the US courts accepted their arguments.³²

To treat the pari passu clause as a dichotomous variable is to equate all bonds containing some version of the clause. This is not merely incomplete; it is misleading. Bonds without priority of payment language may confer entirely different rights than Argentina-style bonds. Indeed, an investor might view a contract with a pari passu clause that lacks priority of payment language, and a contract with no pari passu clause at all, as functionally equivalent. The literature also reveals other systematic variations in the text of pari passu clauses, with important implications for investors’ rights.³³ At minimum, a robust coding scheme should capture whether these variations appear in the text of each bond.

6. CACs

Binary coding also overlooks important differences in CACs. As noted, one important variable is whether the CAC allows for a restructuring vote to be aggregated across different series of bonds. Without this aggregation feature, it is relatively easy for holdout creditors to acquire (at distressed prices) a position large enough to block a restructuring for a given series of bonds. Because commercial databases do not capture this information, policy actors like the IMF had to contract specifically for coding tailored to their needs.

Beyond the question of aggregation, CACs are known to vary in a number of ways not captured by binary coding. The most obvious are the separate voting thresholds for proposals to modify payment terms and for proposals to modify terms unrelated to payment. For instance, recall that Venezuela’s outstanding bond debt includes three separate thresholds for payment modifications (100, 85, and 75 per cent). Moreover, over the past decade or so, some governments have revised their CACs so that the elevated voting threshold—traditionally applicable only to payment-related modifications—applies to a wider range of proposals. Common additions include proposals to modify the bond’s waiver of sovereign immunity and proposals to modify the governing law clause.³⁴ A robust coding scheme should encompass the voting thresholds applicable to a variety of plausible modification proposals.³⁵

7. Other clauses that merit attention

In addition to the refinements discussed above, a robust coding scheme should cover a wide range of clauses presently omitted from commercial databases. This list is extensive, so we will close with just a few examples:

• Grace periods. Sovereign bonds typically allow for a grace period before a government’s failure to pay principal or interest when due constitutes a default. These grace periods (often longer for missed interest payments than for missed principal) vary from bond to bond and can play an important role in determining the consequences of a missed payment. For instance, a short grace period increases the likelihood that a short-term liquidity problem will develop into a wider debt crisis.

• Acceleration rights. It is generally a sign of trouble when a government fails to pay interest on its debt when due. Yet a missed interest payment does not immediately entitle investors to demand their money back. First, there must be a decision to accelerate—ie to declare the loan principal and any accrued interest payable immediately. Most bond contracts collectivize this decision. The contract typically specifies the voting threshold (often 25 per cent) for acceleration and another, higher threshold (often 50 per cent) for rescinding an acceleration. These votes have important implications. For example, acceleration on one bond may trigger the cross-default clause on another, resulting in widespread acceleration across the entire debt stock.

• The use of trust indentures. Many sovereign bonds are issued under a fiscal agency agreement, in which a financial intermediary receives principal and interest payments as the agent of the government. Other bonds rely on a trust indenture, in which the intermediary acts as a fiduciary to bondholders. The difference between loan structures matters. Under a trust indenture, for example, the trustee, not any individual bondholder, is empowered to bring suit against the government, and it will distribute any recovery pro rata to investors. The use of a trust indenture thus provides some deterrent to holdout creditors.

V. CONCLUSION

In section III, we documented significant error rates in the legal coding done by DCM and Bloomberg, two prominent commercial databases. At least until coding improves, it seems imprudent to rely on this data as the basis for important judgements about the sovereign debt markets. Indeed, concerns about data quality may partly explain why academic researchers, IMF staff, and market analysts rely so selectively on Bloomberg and DCM for information about legal terms. This is speculation; we have not seen any explicit reference to doubts about data quality, and some research papers and market analyst reports do, in fact, use commercial databases for information about governing law and CACs. Our findings suggest that even this limited reliance may be imprudent, especially for the presence of CACs.

We hope that this project encourages commercial databases to engage in greater data correction efforts. Perhaps better quality data would encourage more widespread use; certainly we would benefit from improved data in our own research efforts. But we are not optimistic that database owners will expend the resources necessary to increase accuracy. It is hard to explain the combination of high error rates and significant cross-country variance in data accuracy. We speculate that much of the coding has been automated. If so, it is not clear that database owners have financial incentives to identify and correct flaws in coding algorithms.³⁶ It is even less likely they will invest resources in hiring and training coders who can identify legally relevant differences in bond contracts.

Moreover, even if there is some improvement in data quality, we doubt the commercial marketplace will generate sufficiently detailed information about loan contracts. For-profit information providers are primarily responsive to customer demand. Private investors are likely to have a narrower range of interests than academics and policy actors. If private investors are the key customers, we’d expect commercial database providers to overlook a wide range of terms important to other stakeholders. This intuition draws support from the fact that the IMF had to contract separately for coding to capture the use of aggregated CACs and revised pari passu clauses. In any event, robust information about the financial and legal structure of sovereign debt markets is a public good, necessary for sound empirical research and informed policy judgements. There is little reason to think private markets for data will fill this need.

The reality, we suspect, is that high quality data on this market—of the kind that is important to answer big policy questions—requires a coodinated effort of public sector institutions such as the IMF and the Bank for International Settlements. Those institutions have long held debt transparency to be a crucial policy objective, and they already do the public service of providing some crucial information about debt markets to researchers. So far, however, there has been no effort to coordinate disclosure and reporting of legal terms in sovereign bonds.

Footnotes

APPENDIX A

Tamon Asonuma and Christoph Trebesch, ‘Sovereign Debt Restructurings: Preemptive or Post- Default?’ (2015) 14(1) Journal of the European Economic Association 175.

Alfredo Bardozzetti and Davide Dottori, ‘Collective Action Clauses: How Do They Weigh on Sovereigns?’ (2013) 92(2) Journal of International Economics 286.

Michael Bradley and Mitu Gulati, ‘Collective Action Clauses for the Eurozone’ (2014) 18(6) Review of Finance 2045.

Michael Bradley, Elisabeth de Fontenay, Irving de Lira Salvatierra and Mitu Gulati, ‘Pricing Sovereign Debt: Foreign Versus Local Parameters’ (2018) 24(2) European Financial Management Journal 261.

Lee. C. Buchheit and Mitu Gulati, ‘The Gathering Storm: Contingent Liabilities in a Sovereign Debt Restructuring’ in Lee C. Buchheit and Rosa Lastra (eds) Sovereign Debt Management (Oxford University Press, 2014).

Elena Carletti, Paolo Colla, Steven Ongena and Mitu Gulati, ‘Pricing Contract Terms in a Crisis’ (2016) 11 Capital Markets Law Journal 540.

Marcus Chamon, Julian Schumacher and Christoph Trebesch, ‘Foreign Law Bonds: Can They Reduce Borrowing Costs?’ (forthcoming) Journal of International Economics.

Stephen J. Choi, Eric A. Posner and Mitu Gulati, ‘The Dynamics of Contract Evolution’ (2013) 88 NYU Law Review 1.

Stephen J. Choi, Mitu Gulati and Robert E. Scott, ‘Variation in Boilerplate: Rational Design or Random Mutation?’ (forthcoming) American Law & Economics Review.

Paolo Colla, Anna Gelpern and Mitu Gulati, ‘The Puzzle of PDVSA Pricing’ (2016) 12 Capital Markets Law Journal 66.

Juan J. Cruces and Tim R. Samples, ‘Settling Sovereign Debt’s “Trial of the Century”’ (2016) Emory International Law Review 7.

Juan J. Cruces and Christoph Trebesch, ‘Sovereign Defaults: The Price of Haircuts’ (2013) 5 American Economic Journal: Macroeconomics 85.

Elisabeth de Fontenay, Josefin Meyer and Mitu Gulati, ‘The Sovereign Debt Listing Puzzle’ (forthcoming) Oxford Economic Papers <https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2853917>.

Dilip Ratha, Supriyo De and Sergio Kurlat, ‘Does Governing Law Affect Bond Spreads?’ (forthcoming) Emerging Markets Review.

Julian Schumacher, Henrik Enderlein and Christoph Trebesch, ‘What Explains Sovereign Debt Litigation?’ (2015) 58 Journal of Law and Economics 585.

Christoph Groβe Steffen and Julian Schumacher, ‘Debt Restructuring in the Euro Area: How can Sovereign Debt be Restructured More Effectively?’ (2014) 4 DIW Economic Bulletin 19.

W. Mark C. Weidemaier, ‘Sovereign Immunity and Sovereign Debt’ (2014) Illinois Law Review 67.

W. Mark C. Weidemaier and Mitu Gulati, ‘Market Practice and the Evolution of Sovereign Immunity’ (forthcoming) Law and Social Inquiry.

W. Mark C. Weidemaier, Mitu Gulati and Robert E. Scott, ‘Origin Myths, Contracts and the Hunt for Pari Passu’ (2013) 38 Law and Social Inquiry 72.

Jeromin Zettelmeyer, Christoph Trebesch and Mitu Gulati, ‘The Greek Debt Restructuring: An Autopsy’ (2013) 28(75) Economic Policy 513.

APPENDIX B

International Monetary Fund, Strengthening the Contractual Framework to Address Collective Action Problems in Sovereign Debt Restructuring, 1 October 2014.

International Monetary Fund, Progress Report on Inclusion of Enhanced Contractual Provisions in International Sovereign Bond Contracts, 18 September 2015.

International Monetary Fund, Second Progress Report On Inclusion of Enhanced Contractual Provisions in International Sovereign Bond Contracts, 27 December 2016.

International Monetary Fund, Third Progress Report on Inclusion of Enhanced Contractual Provisions in International Sovereign Bond Contracts, 15 December 2017.

International Monetary Fund, The Fund's Lending Framework and Sovereign Debt—Further Considerations, 9 April 2015.

APPENDIX C

Bank of America Merrill Lynch, Surprise Announcement: Maduro Announces PDVSA/Ven Debt Restructuring, 3 November 2017.

Citi Research, Venezuela Credit Strategy Review, 8 August 2017.

Deutsche Bank Markets Research, Venezuela: Preparing for the End Game, 14 July 2017.

J.P. Morgan, Venezuela and PDVSA Debt: A Guide, 30 March 2015.

Morgan Stanley Research, Venezuela Markets and Strategy: Not Getting any Easier, 24 January 2017.

Nomura Capital, Venezuela—Challenge Ahead, September 2017.

Nomura Capital, Venezuela—How the Crisis Evolves, March 2018.

Reorg Research, Investors Gear Up for Grueling Venezuela Restructuring as Maduro Regime Shows Willingness to Engage with Creditors Despite Sanctions, 3 November 2017.

Torino Capital, Venezuela Red Book, Q1 2018.

Author notes