BT Sovereign Signal

Sovereign investors usually look at countries through market data, credit views and country data. Those lenses overlap, but they are not the same.

A country can trade at a high spread because fundamentals are weak. It can also trade wide because of liquidity, investor positioning, recent restructuring history, election risk or a global selloff.

That is all BT-SFS tries to answer. Market comparison is a later use case, not part of score construction.

These rules keep BT-SFS open, auditable and separate from the market prices it is later compared against.

BT-SFS uses public data and a published, source-linked default and restructuring registry where public datasets lag recent events.

All inputs are public data. IMF data are used under the IMF's terms, with attribution ("Source: International Monetary Fund", with the database named), and are materially transformed: BT-SFS publishes scores built from them, not the IMF series themselves. World Bank data carry the Bank's attribution.

The score blends source dates. Current-year IMF fields are projections, while WDI and WGI observations lag by a year or more. Every metric row in a release carries its source year or span.

The headline score uses ten metrics in four pillars, plus a separate default and restructuring adjustment.

Fiscal measures the debt stock, its cost, the primary balance and the part of debt owed in foreign currency. External is reserves in months of imports. Macro covers income (GDP per capita), growth (a centred 3-year mean) and growth volatility. Institutions combines governance (the mean of WGI Government Effectiveness and Regulatory Quality on the absolute 0–100 scale) with a 30-year inflation track record, which captures what monetary credibility delivers in hard data.

Shown as diagnostics, not scored: current account / GDP, IMF reserve adequacy and reserves / short-term debt, economic scale, debt / revenue, central bank independence, WGI Rule of Law, Control of Corruption and Political Stability, the FX share of public debt, gross interest / revenue and total external debt / GNI.

Current account, economic scale and debt / revenue left the score after testing against sovereign distress: the current account showed no signal, scale a weak one, and debt / revenue duplicated debt / GDP (see validation).

Every metric is converted into an oriented score where 100 is better and 0 is worse, measured against fixed goalposts rather than against the other countries in the release.

The reference pool is every universe country-year from 2010 to 2019 (49 countries; 490 observations per metric, 461 for FX debt and 464 for reserves). Values beyond a goalpost are winsorised to 0 or 100. Because the goalposts are fixed, scores are comparable across years and a broad deterioration shows up as lower scores for everyone, not as a reshuffle. The goalposts are frozen in the release's specification lock and reviewed every five years or with a methodology change; the within-release rank is still published as a display rank.

GDP per capita and growth volatility are scored in logs, so each doubling counts the same; the table converts their goalposts back to dollars and percentage points. The inflation record is a stress index; its goalposts are equivalent to steady inflation of about 2.4% and 35.4% a year.

Inflation track record

The inflation track record replaces the v0.4 single-year inflation bands. It weights 30 years of IMF CPI inflation with a 10-year half-life, so about 57% of the weight falls on the last decade: a bad year fades gradually instead of dropping out of a fixed window on a set date, and genuine disinflation is rewarded gradually. Deflation counts double, preserving the v0.4 view that falling prices are also stress.

Growth window

Real growth is the mean of the IMF series for the prior, current and next year. The window is kept for smoothing, so that a one-off shock such as a drought does not define a country; it is not there for predictive gain. Backtests use the growth forecast as published in each year's October WEO: a centred window evaluated on realised data would be look-ahead, and on as-published data the window scores the same as a single year. Years without a published vintage use the trailing two years, flagged.

The base score is the weighted arithmetic mean of the ten metric scores. An arithmetic mean is compensatory: strength in one metric can offset weakness in another. A pillar-level geometric mean, which limits that offset, is computed as a robustness check but does not set the headline score.

The default and restructuring adjustment is separate because recent default history is not just another macro variable. It captures willingness to pay, market-access track record and restructuring complexity. It is source-linked, time-decayed, capped and visible in every release.

The penalty holds at full size for two years, then halves every five. A fresh high-severity default already sits at the −12 cap; about 3½ rating notches. No country loses more than that for its default record, however many events it has: countries with several recent events already score low on post-default fundamentals, and a deeper cap would count that weakness twice.

The overlay interacts with the metrics by design: a restructuring can improve measured debt ratios while the penalty offsets it, and the offset fades on the published schedule rather than on new behaviour.

Agency ratings and market spreads are useful. They are also exactly the things we want to compare the score against.

Metrics are chosen and weighted against sovereign distress (defaults and restructurings), never against spreads. Choosing metrics by how well they explain spreads would bring the market back into the score by the back door, and the comparison with spreads would then be circular.

One practical use is a chart comparing BT-SFS with USD sovereign bond spreads. That chart is a product and research surface. It is not the methodology itself.

BT-SFS reflects several inputs at once: academic work on ratings, spreads and debt distress; public rating-agency and IMF methodology; the composite-indicator literature; open-data constraints; and our own requirement that fundamentals stay separate from market pricing. No single external model is the template.

Cantor and Packer show that macro, income, inflation, external and default-history variables explain much of sovereign rating variation. Eichengreen and Mody, Hilscher and Nosbusch, Longstaff et al. and others show that spreads reflect both country fundamentals and global risk conditions, and Dell'Erba, Hausmann and Panizza find that debt is priced mainly where it is owed in foreign currency. Mauro, Sussman and Yafeh add a longer historical comparison across the 1870–1913 and modern globalisation periods.

That boundary is important. BT-SFS is an inspectable fundamentals score that can be compared with ratings, spreads and default outcomes. It is not a rating model, a default model or a pricing model.

BT-SFS measures observable fiscal, external, macro, institutional and default-history fundamentals. It does not fully capture policy credibility beyond the inflation record, fiscal accounting quality, FX-regime fragility, political transition risk or market technicals. Reserves in months of imports cover trade needs, not rollover risk against short-term external debt, which is shown as a diagnostic. Those forces may appear as gaps between the fundamentals score and market spreads.

BT-SFS is not a sovereign rating, default-probability model, recovery model, investment recommendation or fair-value model for spreads.

Historical backcast policy

Historical series are current-vintage backcasts: they apply the current methodology and today's data to past years, with growth taken from the IMF forecast as published at the time and FX debt from QPSD history, then the World Bank external public-debt proxy, then the latest known share carried back (each flagged). They do not show what the model would have known in real time. Public charts draw complete years as solid lines and leave missing years as gaps rather than inventing values.

The paper explains how the score is built; the validation appendix asks whether it behaves sensibly against observable outcomes. For v0.5 the acceptance criteria were written down before the composite score was ever run against the event panel.

The test was pre-registered in commit dc2744e72; build amendments were logged before the run in 0af6cd3d3 and 6c9952ab5; the scoring pipeline that ran it is 0018b29d0. One amendment came after the run: A9 (reserves not applicable for Panama, which has no central bank) does not move the panel, and the test was re-run on the rebuilt release. v0.4 and v0.5 were recomputed on the same panel: the corrected 20-year registry, a 3-year horizon, and rows excluded from each event's onset until two years after it was resolved. The panel has 361 country-years, 22 event rows, 7 event countries and 8 episodes. v0.5 passed all four criteria.

How to read the table: the AUC is the chance that, of one country-year followed by a default and one that was not, the score ranks the one heading into default lower (0.5 is a coin toss, 1.0 is perfect separation). The weakest quartile is the lowest-scoring quarter of all country-years; the lift is how many times more often defaults follow a year there than anywhere else. The ranges in brackets show how much the result would move with a different mix of countries.

Caveats. This is an in-sample comparison on a small panel: seven event countries, with rows that are not independent. The metrics were screened against distress on an event panel that overlaps this one, so part of the gain reflects that selection. The backcast uses current-vintage data, and the broader stress outcome is not yet assembled. The result supports v0.5 as a better-built fundamentals score; it does not establish forecast skill.

The validation appendix is published alongside this paper: BT-SFS validation appendix.

Each release publishes this paper, the validation appendix, the score specification lock with its goalposts, the country universe, source-traceable inputs with source years and flags, metric, pillar and base scores, the default adjustment, final scores, the event registry with source links, the FX currency table, known source failures, and the version history.

The universe is the USD sovereign issuers in a benchmark EM sovereign index, plus countries carried forward while BondTerminal carries a live USD bond. Agency ratings appear only as reference fields, and nothing in the release is investment advice.

The data package is public at bondterminal.com/bt-sfs-paper/data: scores, every metric input, the goalposts and their reference pool, the event registry, the FX currency table, the specification lock and the validation outputs, each with a SHA-256 hash. The scoring code is not public, so the package is built for checking scores by hand; the validation appendix walks through one country. The raw IMF series cached with the release are not redistributed until the IMF grants permission.

This is an open project, not a closed claim of authority. Good feedback is concrete enough to test.

The standard for change is simple: does the proposal make the model more transparent, more reproducible, more stable or more useful, without pretending to be more precise than the data allow? Changes are tested against sovereign distress and logged in the version history.

The exhibits below show the v0.5 release. They come after the methodology on purpose: the table, the spread view, the history view and the profile matrix are evidence and reader aids, not inputs into the score.

Release bt-sfs-v0-5-2026-09-27, IMF target year 2026, market date 2026-09-25. The spread is one selected USD benchmark bond's yield to maturity minus the 5-year US Treasury yield, with the bond chosen for modified duration near five years; notes flag weak proxies and sparse curves. Countries are ordered by rank.

Metric Profile Matrix

Each cell shows the metric score against its fixed goalposts, then the raw value and its source year. Higher score is better for every metric. A "p" marks IMF projection years; spans show multi-year windows.

1. IMF DataMapper API help
2. IMF DataMapper general government debt/GDP example
3. IMF, World Economic Outlook historical forecasts database
4. World Bank API developer documentation
5. World Bank WDI reserves months example
6. World Bank, Quarterly Public Sector Debt (QPSD)
7. Eurostat, government debt by currency (gov_10dd_dcur)
8. World Bank Worldwide Governance Indicators
9. World Bank, WGI 2025 methodology revision and documentation
10. Bank of Canada / Bank of England sovereign default database 2025 update
11. Bank of Canada technical report on the sovereign default database
12. IMF (2021), Review of the Debt Sustainability Framework for Market Access Countries
13. IMF (2022), Staff Guidance Note on the Sovereign Risk and Debt Sustainability Framework (SRDSF)
14. IMF and World Bank (2018), Guidance Note on the Debt Sustainability Framework for Low-Income Countries
15. IMF (2016), Guidance Note on the Assessment of Reserve Adequacy
16. VanEck, the investment case for emerging-markets debt
17. Cantor and Packer, determinants and impact of sovereign credit ratings
18. Eichengreen and Mody, what explains changing spreads on emerging-market debt
19. Mauro, Sussman and Yafeh, emerging markets and financial globalization
20. Afonso, Gomes and Rother, sovereign rating determinants
21. Hilscher and Nosbusch, determinants of sovereign risk
22. Longstaff, Pan, Pedersen and Singleton, how sovereign is sovereign credit risk?
23. Dell'Erba, Hausmann and Panizza, debt levels, debt composition and sovereign spreads
24. Cruces and Trebesch, sovereign defaults: the price of haircuts
25. Manasse, Roubini and Schimmelpfennig, predicting sovereign debt crises
26. Kraay and Nehru, when is external debt sustainable?
27. Hellwig (2021), predicting fiscal crises: a machine learning approach
28. IMF technical note on vulnerability indicators and machine learning
29. Reinhart, Rogoff and Savastano, debt intolerance
30. OECD and JRC (2008), Handbook on Constructing Composite Indicators
31. Paruolo, Saisana and Saltelli (2013), ratings and rankings: voodoo or science?
32. LSEG sovereign risk monitor methodology
33. S&P Global Ratings sovereign rating methodology
34. Fitch Ratings sovereign rating criteria (2024)
35. Fitch Sovereign Rating Model (SRM) coefficients, in the 2024 criteria
36. Moody's sovereign methodology overview
37. Scope Ratings, sovereign rating methodology (2026)