The 37 percent average haircut reproduces almost exactly; the price of haircuts itself is untestable on this estate

The paper, and what is rerun here

The original is Juan J. Cruces and Christoph Trebesch, “Sovereign Defaults: The Price of Haircuts”, American Economic Journal: Macroeconomics5(3), 2013. The paper constructed investor-loss estimates (“haircuts”, in the present-value metric of Sturzenegger and Zettelmeyer, here SZ) for sovereign restructurings with private external creditors over 1970-2010, reported a mean of about 37 percent on 180 cases, and then estimated its headline results: restructurings with larger haircuts are followed by longer exclusion from capital markets and by higher post-restructuring bond spreads. The 1970-2010 window, the SZ haircut as the headline concept, and dating deals by completion year are the paper’s choices, kept here.

This note reruns what the estate can support: the haircut distribution on the CT-lineage file (187 rows, of which 9 are flagged as 2014-update additions and excluded, leaving 178), and one extension on the estate’s longer file, which runs 1823-2024: did haircuts change after 2010? One robustness variant was pre-declared and no other: restricting the post-2010 window to deals whose full_repudiationflag is explicitly 0. The paper’s regressions are not attempted, for reasons laid out below, and no result on exclusion or spreads is claimed in either direction.

The distribution reproduces

On the 178 original-case restructurings completed 1978-2010 across 70 economies, the mean SZ haircut is 36.70% against the paper’s reported ~37% on 180 cases: the headline moment survives this vintage with a 2-case gap in the sample. The distribution is right-skewed (median 31.70%, sd 27.18) and runs from -9.8% to 97.0%: 6 deals have negative haircuts, meaning investors came out ahead in present value, and they are kept, not clipped. Weighting by the nominal USD amount restructured pulls the mean down to 31.07%.

The file carries three haircut concepts and they are far apart. The face-value reduction has a median of 0.00%: 122 of the 178 deals cut no face value at all, so whatever present value those deals took from investors came through terms other than principal reduction.

Source: Cruces and Trebesch (2013), "Sovereign Defaults: The Price of Haircuts", AEJ: Macroeconomics 5(3); dataset as updated 2014 (haircuts_cruces_trebesch_2014.parquet) All three concepts on the same 178 restructurings. Volume-weighted mean SZ haircut 31.07% (weights: nominal USD millions restructured, n = 178).

After 2010: smaller haircuts, unless you weight by dollars

The paper’s window closed in 2010. The estate’s extended file adds the deals since, so the natural extension question is whether haircuts got bigger. Run within the one file (its 1823-2024 span cut to 1970 onwards, to stay comparable with the paper’s window): the answer on the unweighted margin is no. Deals completed 1970-2010 average 38.05% (n = 170); the 23 deals of 2012-2024 average 26.58%. Welch’s t = 2.126 (p = 0.042); the Mann-Whitney test, less impressed by tails, gives p = 0.077. Marginal at best, and resting on 23 post-2010 observations.

Dollar-weighting flips the sign: 30.10% for 1970-2010 against 54.54% after. Most of that flip is one deal. Greece 2012 restructured USD 261.4bn at a 64.6% haircut, 66.8% of all post-2010 volume in the file; drop it and the weighted post-2010 mean falls from 54.54% to 34.32%, back within 4.22 points of the earlier window’s 30.10% (unweighted 24.85%, n = 22). Both margins are shown because reporting either alone would mislead: the typical post-2010 deal cut less than the typical earlier one, and the typical post-2010 dollar was cut more, because most post-2010 dollars were Greek.

Source: Extended haircut dataset (haircuts.parquet); its source field cites Meyer, Reinhart and Trebesch (2022), Cruces and Trebesch (2013), Asonuma and Trebesch (2016), and consent solicitation documents and ratings reports Both windows from the extended file only; its 1970-2010 window (n = 170) does not match the CT file's 178 case-for-case, so no cross-file test is run. Weighted means use nominal USD millions restructured.

The pre-declared variant: the file’s full_repudiation flag marks 9 deals, all at exactly 100.0% and all completed 1823-1960, before either window. Restricting the post-2010 sample to deals explicitly flagged 0 therefore removes no repudiation; what it removes is the one deal whose flag is uncoded, Ukraine 2016 (a 19.2% haircut on USD 511m), which a naive != 1 filter drops silently. The result: post-2010 mean 26.91% (n = 22), Welch p = 0.056.

Source: Extended haircut dataset (haircuts.parquet); its source field cites Meyer, Reinhart and Trebesch (2022), Cruces and Trebesch (2013), Asonuma and Trebesch (2016), and consent solicitation documents and ratings reports All 23 deals completed after 2010, in completion order. Country-year pairs can repeat as genuine separate deals; across the whole file 4 pairs carry two deals each (Poland 1982, Pakistan 1999, Russia 2000, Ukraine 2016).

The regressions this estate cannot run

The paper’s contribution is not the mean; it is the price. Its Cox duration models need the date each defaulter regained market access, and its spread regressions need a post-restructuring sovereign bond spread panel. Neither exists here. Across the 34 distinct column names of the estate’s three haircut files, a scan for access, spread or embi returns 0 hits; a scan of the estate’s 167 other parquet tables for embi, spread or yield in the name returns 1: ust_yield_curve, the US Treasury curve, not a sovereign spread panel.

The nearest-looking object, haircuts_default_spells.parquet, carries default-spell durations (year_start to year_end). That measures how long a country stayed in default; the paper’s dependent variable measures how long, after settling, the country stayed shut out of capital markets. Substituting one for the other would test a different hypothesis under the paper’s name, so this note stops at the distribution and the post-2010 comparison.

Two lineages, and neither is the paper’s dataset

The CT-lineage file holds 187 restructurings: 178 flagged as original cases plus 9 added in the 2014 update. The paper reports 180 cases; this file’s original-case count is 178. The 2-case gap is unreconciled here, and since the mean on 178 cases lands within a third of a point of the paper’s reported ~37%, it reads as case consolidation or omission in this file’s build rather than a units problem.

The extended file is a different lineage: 329 deals over 1823-2024, whose source field cites Meyer, Reinhart and Trebesch (2022), Cruces and Trebesch (2013), Asonuma and Trebesch (2016), and consent solicitation documents and ratings reports, across 10 raw source strings (one of them a typo variant, Cruces and Trebesch (2013)) with a doubled parenthesis, which is why nothing here groups by raw source). Its 1970-2010 window holds 170 deals where the CT file holds 178: the lineages disagree case-for-case even in their overlap, so the pre/post-2010 comparison runs entirely within the extended file and the distribution moments entirely within the CT file.

What this rerun cannot tell you

• Nothing about exclusion or spreads. The paper’s headline regressions were not run, so this note neither confirms nor contradicts the “price of haircuts” itself. What reproduces is the distribution the regressions were built on.
• The post-2010 comparison rests on 23 deals. The two tests disagree at the 5 percent line (Welch p = 0.042, Mann-Whitney p = 0.077), so “haircuts got smaller” is a marginal finding, not a settled one. A single future mega-restructuring could move the post-2010 column materially.
• The weights are nominal dollars across five decades. The volume-weighted means use debt_restructured_usd_mas reported, mixing 1980s and 2020s price levels; the extended file’s real-2020 column was not used, to keep one weighting convention throughout. The Greece decomposition is the honest way to read the weighted levels.
• The 178-versus-180 gap is unexplained. No case-by-case reconciliation against the paper’s appendix was performed, and no figure from the paper beyond its reported mean of about 37 percent, its case count of 180, and its 1970-2010 window is quoted or tested.
• The absence scans are name scans. The reaccess and spread variables were declared missing after scanning column names of the three haircut files and the names of 167 estate tables. A spread panel hiding under an unrelated name would not have been found.
• The extended file’s deals are not all built the same way. Its source field cites Asonuma and Trebesch (2016) and consent solicitation documents and ratings reports for recent deals rather than the CT paper’s procedure, so part of any pre/post difference could be construction, not economics.

Built from scripts/build/build_research_haircut_size.py reading haircuts_cruces_trebesch_2014.parquet and haircuts.parquet; every computed value on this page is interpolated from its output. Haircut percentages are on a 0-100 scale (negatives included); restructured amounts are nominal USD millions as carried by the files.