Sovereign Credit Rating Explorer

§ Methodology

This page documents the data construction, scoring conventions, and relative rating methods (M1–M4) used throughout the Explorer. For term definitions, see the Glossary.

1 Raw Rating Data

Sources and Coverage

Credit ratings are collected from three agencies: Standard & Poor's (S&P), Moody's, and DBRS Morningstar. The dataset covers 151 sovereign entities from 1949 to 2026, with 2,178 rating events in total. Coverage is uneven across agencies: S&P and Moody's rate the broadest set of countries; DBRS focuses on a smaller subset, skewed toward higher-income economies.

Numeric Scoring (0–60 Scale)

Each agency uses a letter-grade scale (e.g. AAA, AA+, AA, …, D). These are converted to a uniform numeric scale from 0 (worst) to 60 (best) for comparability across agencies. The approximate mapping at key grades:

2 Panel Construction

Annual Forward-Fill

Rating events are dated to the day but are sparse — a country may go years between rating actions. To construct a balanced annual panel, each country-agency rating is forward-filled: once a rating is assigned, it is carried forward to all subsequent years until a new rating event supersedes it. This reflects the convention that an outstanding rating remains in effect until it is explicitly changed.

Composite Score

For each country-year, the composite score is the simple arithmetic mean of all available agency scores for that country in that year. Countries rated by only one agency receive that agency's score as their composite; countries rated by all three receive the mean of three scores. The composite is not computed for country-years with no outstanding rating from any agency.

Macro Covariates

GDP (current USD), population, and FDI net inflows (% of GDP) are sourced from the World Bank Development Indicators and merged by ISO3 country code. GDP per capita is derived as GDP / population. Log GDP per capita (base 10) is used in regressions to reduce skewness. Coverage: GDP ≈ 94%, FDI ≈ 92%, GDP per capita ≈ 94% of panel rows. Missing values: Guinea and Taiwan (World Bank does not publish Taiwan's national accounts).

3 Peer Groups

Countries are assigned to 59 peer groups based on geographic, political, and economic criteria. Groups range from broad (World, 150 countries) to narrow (GCC, 4 countries). The complete list includes:

• G-groups: G7, G8, G10, G20, G24, G15
• European clubs: EU, Eurozone, Schengen, NATO, Nordic, Baltic, CIS, and sub-regional groupings
• Asian sub-regions: ASEAN/Southeast Asia, SAARC/South Asia, East Asia, GCC, Arab League, and others
• Americas: Latin America, MERCOSUR, Andean, Caribbean, etc.
• African sub-regions: North, West, East, Southern, Central Africa
• World Bank income tiers: High Income, Upper Middle, Lower Middle, Low Income
• World: all 150 rated countries, used as the global benchmark

4 Relative Rating Methods (M1–M4)

Absolute scores measure a country's credit quality on a fixed scale. Relative ratings measure a country's standing within its peer group, which can change even when the absolute score is unchanged — for example, if the group mean deteriorates. Four methods are implemented, each capturing a different dimension of relative standing.

Leave-One-Out (LOO) Group Mean

All relative methods use a leave-one-out (LOO) mean as the group benchmark. The LOO mean for country i in group g at year t is the simple mean of all other members' composite scores:

Using the LOO mean prevents a country's own score from influencing its benchmark, which would mechanically compress relative ratings toward 1 and bias estimates of relative standing, particularly in small groups.

M1 — Current Relative Rating

M1 scales a country's score against the current-year LOO mean. The score is squared before dividing to amplify differences at the upper tail of the distribution (where most rated countries cluster):

M1 > LOO mean indicates above-average standing; M1 < LOO mean indicates below-average standing. M1 captures both genuine rating changes and passive drift: a country holding a fixed absolute score will see its M1 rise as the group mean falls, and vice versa.

M2 — Within-Year Percentile Rank (Raw Count)

M2 is the ordinal rank of country i within its peer group in year t, expressed as a raw integer (e.g. 47 out of 150). A higher rank means a higher composite score relative to peers.

M2 is non-parametric and robust to the shape of the score distribution, but it is bounded by group size and does not distinguish between large and small score gaps.

M3 — Lagged Relative Rating

M3 is identical to M1 except the benchmark uses the prior year's LOO mean rather than the current year's:

M3 is most useful for studying the speed of adjustment in relative ratings. At the onset of a crisis, M3 is systematically more optimistic than M1: because last year's LOO mean has not yet reflected the group's deterioration, the denominator is higher, making the ratio smaller and the relative rating appear better than M1 suggests. Conversely, during recovery M3 lags behind M1 as the old (poor) group mean persists. M3 is not defined for the first year a country-group series begins (no lag available).

M4 — Within-Year Percentile (0–1 Scale)

M4 is the percentile rank of country i within its peer group, normalised to the unit interval:

M4 is more comparable across groups of different sizes than M2. A value of 0.5 means the country is at the median of its group. M4 saturates near 1 at the upper tail — a country at the absolute maximum score (60) will reach M4 ≈ 0.97 rather than exactly 1.0 because at least one other country (itself, excluded) is assumed to exist.

5 Comparing M1–M4: Guidance

6 Gini Coefficient (Distributional Inequality)

For each peer group in each year, a Gini coefficient is computed over member composite scores to measure within-group inequality in credit quality. A Gini of 0 means all members have identical scores; a Gini of 1 means one member holds all the score mass.

The World Gini (group = "World") rose monotonically from 0.232 in 2000 to 0.321 in 2023, indicating that global credit inequality has grown approximately 38% over the past 25 years — driven by persistent divergence between high-income stable sovereigns and volatile emerging/frontier markets.

7 Panel Regression Framework

The main regression dataset (paper_panel.csv) supports the following specifications. All panel regressions use clustered standard errors at the country level.

The stark contrast between Models A and D (R²(within) ≈ 0 vs R²(overall) = 0.90) is the central empirical result: credit scores are almost entirely determined by structural country characteristics rather than annual macroeconomic fluctuations. The between-country component — richer countries get better ratings — dwarfs any within-country year-to-year variation.

8 Ordered Model for Rating Transitions

A separate ordered logit and ordered probit are estimated on the sample of annual rating transitions (N = 2,922 country-year pairs), where the dependent variable takes three values:

• −1: downgrade (composite score fell from prior year) — 13% of observations
• 0: stable (no change) — 70% of observations
• +1: upgrade (composite score rose from prior year) — 17% of observations

Regressors: GDP growth (prior year), log GDP per capita, FDI (% GDP), and lagged composite score. Key results: GDP growth and log GDPpc are positively associated with upgrades; the lagged score is negatively associated (mean reversion — already high-rated countries have less room to rise); FDI is insignificant.

Last updated April 2026. Data: S&P, Moody's, DBRS; World Bank Development Indicators. Analysis: Python (pandas, linearmodels, statsmodels).