Sovereign haircuts / methodology
Methodology, the measures, and what this data cannot tell you
The tables
| Table | Rows | iso3 codes | Span | What it holds |
|---|---|---|---|---|
| haircuts | 329 | 94 | 1823-2024 | Concluded restructurings and the haircut creditors took |
| haircuts_cruces_trebesch_2014 | 187 | 72 | 1978-2013 | The academic haircut file: three measures, the exit yield, deal structure |
| haircuts_default_spells | 206 | 93 | 1823-2024 | Default episodes: start, end, duration |
| crag | 46,931 | 166 | 1960-2024 | BoC-BoE: stock of debt in default by creditor class, including ongoing defaults |
| china_debt_restructurings | 55 | 29 | 2000-2020 | China's restructurings of its overseas loans. No haircut column |
Three measures, and how the data says which is which
A sovereign haircut can be measured in at least three ways, and the three do not give the same answer for the same deal. This layer keeps them in separate columns and never averages across them.
haircut_pct)The new instruments, discounted, against the old claim, discounted at the same exit yield. This is the measure carried in both haircut files. Median across the 329 episodes: 37.0%.
haircut_market_pct)The new instruments, discounted, against the FACE value of the old claim. It exists only in the Cruces-Trebesch file. Median across its 187 episodes: 40.3%, against 32.5% for the present-value haircut on the same deals.
haircut_facevalue_pct)Principal written off, and nothing else. Median across the 329 episodes: 0.0%, because 165 of them wrote off no principal at all.
The identification is not a guess. If the old claim has already fallen due, its present value is its face value, so discounting it can change nothing and the present-value haircut must equal the market haircut. In the Cruces-Trebesch file all 94 episodes flagged all-debt-fallen-due have the two measures equal, and only 1 of the other 93 does. The exit yield doing the discounting is itself an input, and it is not a narrow one: it runs from 7.9% to 41.2% across the 166 episodes that carry one, with a median of 14.8%. 21 episodes carry no exit yield at all.
Do the two haircut files mean the same thing?
160 episodes appear in both files under an iso3-year that is unique on both sides, which is the only way to match them: iso3 and year together are not a key in either file. On those 160, 158 agree on the present-value haircut to within 0.05 points and 160 agree on the face-value reduction. The two files therefore use the same definition for the columns they share, and the competing measures live inside the Cruces-Trebesch file rather than between the files.
The 2 matched episodes where the present-value haircut differs are Yugoslavia 1984 (-1.3% in the long-run file, -7.5% in Cruces-Trebesch) and Russia 1997 (0.0% in the long-run file, 26.2% in Cruces-Trebesch). Both numbers are printed on the country page and neither is preferred here.
Where each haircut comes from
| Source study | Episodes | Span |
|---|---|---|
| Meyer, Reinhart & Trebesch (2022) | 136 | 1823-1993 |
| Cruces & Trebesch (2013) | 174 | 1940-2013 |
| Authors' calculations from deal documents | 14 | 2015-2024 |
| Asonuma & Trebesch (2016) | 5 | 2015-2023 |
The source column is free text and carries a typo variant of one study's name, so the families above are matched on a prefix rather than on equality, and the build throws if an episode falls outside all four prefixes.
Coverage: the two centuries are not one record
1950-1979 holds 19 restructurings across 17 countries; 1980-1999 holds 142 across 61. The index page reports a pooled median over 1823-2024 (37.0%) and a median within each era, and the two should be read together: 142 of the 329 episodes fall in 1980-1999, the single densest era, so the pooled figure leans on it. The era boundaries used here (1900, 1950, 1980, 2000) are an analytic choice; the first and last labels are read from the data, and the build throws if any era bucket empties.
The long-run file carries no discount-rate column at all. The exit yield exists only in Cruces-Trebesch, whose earliest episode settled in 1978. For everything before that, the haircut is in the file but the rate that produced it is not, and this layer has no way to re-derive it.
Default spells
206 spells across 93 countries. Duration counts both endpoints in all 206 of them, so a default opened and settled in one year has duration 1. A spell's end year is the year of its last restructuring in 203 of the 206 spells and its start year is the earliest default year of those restructurings in 205 of 206, so the obvious reading of the endpoints is close but not exact.
The spells table also carries haircut_cumulative_pct. It is not published anywhere in this layer. On the 142 spells that contain exactly one restructuring it should equal that restructuring's own haircut, and it does in only 87 of them: it differs in 55. Whatever weighting the column applies is not documented in the file, so it is left out rather than presented as a haircut.
1 default-spell id is referenced by an episode but absent from the spells table: ALB_1991-1995 (Albania). That is why the spells table covers 93 of the 94 countries with a restructuring.
What this data cannot tell you
A haircut needs a concluded deal, so an unfinished default cannot appear. In 2024 the BoC-BoE default database records 104 governments in default on $542.63B; 44 of those 104 countries have no concluded restructuring anywhere in 1823-2024, and only 9 restructurings in the file concluded from 2020 onward. The sample is a sample of defaults that got resolved.
The haircut is a loss on the debt actually restructured. Neither file carries the country's total public debt, so the share of all creditors' claims that was written down cannot be computed from these tables. 329 episodes have a debt-restructured amount; none have a denominator.
The haircut is computed on the exchange terms for a holder who tendered. A creditor who bought the defaulted bond at a distressed price, or sold before the exchange, did not experience this number, and the price they paid is not in the data.
China's 55 restructurings, in 29 countries from 2000 to 2020, carry no haircut column of any kind. What is recorded is the type (49 rescheduling only, 4 face-value reduction, 2 not recorded) and an interest-rate flag that is set on 10 deals and absent, never zero, on the rest, so absence is not evidence of no rate cut. The present-value loss on a maturity extension, which is what a rescheduling is, is unmeasured for every one of them.
The long-run file has no discount-rate column at all, so for every episode outside Cruces-Trebesch the haircut is in the data but the rate that produced it is not. The rate can be inspected only for the 166 Cruces-Trebesch episodes that record one, the earliest of which settled in 1978. Nothing in this layer can re-derive a haircut from a market price.
Which countries get a page
A country gets a page if and only if it has at least one concluded restructuring in the haircut file: 94 of them. The page list is built from the table, so a country with no episode has no page rather than an empty one.
12 countries appear in China’s restructuring record with no episode in the haircut file, and so have no page here: Angola, Benin, Djibouti, Kazakhstan, Kyrgyz Republic, Maldives, Myanmar, Nepal, Sri Lanka, Tajikistan, Tonga, Vanuatu.
One iso3 code carries more than one country name in the file (Serbia, whose 19th- and 20th-century episodes are labelled Yugoslavia and whose 2004 deal is labelled Serbia and Montenegro). The page title takes the name attached to the most recent episode. The Cruces-Trebesch file additionally files the Yugoslav deals under the iso3 code YUG, which has no counterpart in the long-run file; those deals are shown on the Serbia page and labelled as such. The build throws if a refresh introduces any other iso3 code in the Cruces-Trebesch file with no counterpart, so no episode can go silently pageless.