What Predicts Restructuring of Sovereign Loans Held by the InterGalactic Banking Clan: A Cause-Specific Hazard Analysis of 342 Loan Files, 40–25 BBY

1. Introduction

The InterGalactic Banking Clan, seated on Muunilinst, is one of the great financial powers of the galaxy, and Muuns hold a prominent place among its ranks. A large part of its business is lending to governments: planetary administrations, and the sector governments that stand above groups of member worlds. Enforcement against a world's government is as much a matter of negotiation as of law, so loans to sovereign borrowers end in more ways than repayment or loss. Many are renegotiated, with the term extended, the annual service reduced or the instrument exchanged for a new one. This paper takes the vantage of 24 BBY and asks which features of a sovereign loan, known when its file is opened, predict that it will later be restructured.

Burden ratios of the general kind we use, which set scheduled service against revenue, have a long place in public finance (Ombrell, 30 BBY). Dependence on one export has been linked to fiscal stress in Outer Rim economies (Pethil & Ansaro, 28 BBY), and a concentration index for trade lanes has been proposed for planetary exporters (Drovan, 29 BBY). The Senate's research office has surveyed the recurring obligations that member worlds owe to larger powers (Galactic Senate Legislative Research Office, 27 BBY), though its survey does not follow those worlds' borrowing. To our knowledge no study has set these factors against a lender's own record of outcomes.

A lender's register contains files only for loans that were granted, and the applications that were declined leave no trace in it (Mendrik, 33 BBY). Whatever we estimate is therefore conditional on approval. The question is therefore what predicts restructuring among borrowers the Clan chose to lend to.

We had three hypotheses, fixed before estimation. A heavier debt-service burden, dependence on a single trade lane or export, and a tribute obligation to a larger power would each raise the hazard of restructuring. Three lender-side features, syndication, collateral coverage and loan size, were adjustment variables without a directional prediction. Restructuring competes with two other endings, repayment as scheduled and default followed by seizure of collateral, so we analyse it as one cause among several.

2. Methods

Setting and vantage. The analysis was carried out in the Economic Analysis Office on Muunilinst and completed in 24 BBY. The source is the sovereign series of the Clan's loan registers (InterGalactic Banking Clan, 25 BBY), an extract of which was closed at the opening of 24 BBY. Durations are counted in standard years from opening to a file's first recorded ending or to the closing date, so a file opened in 25 BBY contributes at most one standard year.

Sample. The series held 391 files opened between 40 and 25 BBY. We excluded 49: 31 with no revenue annex, from which a burden ratio cannot be computed, and 18 below a size floor of one million credits. That left 342 files. The 342 files belong to 165 borrowers, so files are not independent, and the analysis allows for that. Sums are stated in credits, used as a generic unit of account.

Outcomes. Each file was assigned to one of four states at the close of the extract. A file was restructured if its register entry recorded a formal change to the scheduled terms agreed with the Clan, whether an extension of term, a reduction of annual service or an exchange for a new instrument (coding conventions follow Vurrel, 26 BBY). It was defaulted and collateralised if the Clan took the pledged assets or revenue stream after a missed schedule, and repaid as scheduled if the final scheduled payment was recorded. Files with none of these entries were open and are censored at the closing date. Only the first ending is analysed: a loan restructured and later defaulted is counted as restructured, at the date of restructuring.

Predictors. The debt-service burden ratio (DSBR) is our own construction: the scheduled service due in the first full standard year, divided by the borrower's mean annual revenue over the three standard years before opening, both taken from the annex. Revenue is measured before any tribute is paid. Single-lane dependence is coded 1 when the annex attributes at least 60% of export value to one trade lane or one export good; the 60% threshold is ours and was fixed before analysis. Tribute obligation is coded 1 when the annex records a recurring obligation from the borrower to a larger power. The lender-side variables are syndication (co-lenders present, against the Clan as sole lender), collateral coverage (pledged value divided by principal at opening) and loan size, entered as the base-2 logarithm so that its coefficient refers to a doubling. We also adjusted for borrower type (sector against planetary government) and for opening cohort, entered per five standard years later.

Statistical analysis. The primary model is a Cox proportional-hazards model for the cause-specific hazard of restructuring, with defaults, repayments and open files treated as censored at their own dates (Tarrow, 27 BBY). The 102 restructurings gave about 13 events per estimated coefficient. Because the same borrower can hold several files, confidence intervals and p-values use standard errors from a bootstrap that resamples borrowers rather than files (Halvane, 31 BBY), with at least 300 resamples per model, and are Wald-type. The interval for the ratio of the early and late DSBR hazard ratios comes from the same bootstrap, resampling both coefficients together. Cumulative incidence over ten standard years was estimated for all three endings by the Aalen–Johansen method. Proportionality for the DSBR was examined by fitting separate coefficients for the first four standard years and afterwards; the cut point was fixed before fitting and no other was tried. A cause-specific hazard ratio describes the rate among files still open on schedule. It is not an effect on cumulative incidence, which also depends on the competing endings.

Sensitivity analyses for censoring. Open files, most of them opened late in the series, are the central threat. We bounded the ten-year proportion restructured by treating every file with unresolved ten-year status first as never restructured and then as always restructured; refitted the primary model with follow-up truncated at six standard years; refitted it on the 130 files opened between 40 and 34 BBY, those with at least ten standard years of potential observation (a boundary set by the closing date, not by the bands of Table 1); and fitted a logistic model for restructuring within ten standard years on files whose ten-year status was resolved. In a composite analysis, default and collateralisation were counted with restructuring as one distress event.

Exposure-definition check. Because tribute reduces the revenue a borrower keeps, we also replaced the DSBR by a version with revenue net of tribute, so that the burden measure and the tribute indicator do not overlap.

Coding reliability. A second analyst coded single-lane dependence, tribute obligation and outcome state independently on a random sample of 120 files. The DSBR was abstracted a second time for 60 of those files. Agreement is Cohen's kappa with Wald intervals for the categorical codes, and a two-way consistency intraclass correlation for the DSBR.

3. Results

Of the 342 files, 122 (35.7%) were loans to sector governments and 220 (64.3%) to planetary governments. The mean DSBR was 0.30 (SD 0.09; range 0.05–0.57). Single-lane dependence was recorded for 160 files (46.8%), a tribute obligation for 74 (21.6%) and syndication for 158 (46.2%). Mean collateral coverage was 1.20 (SD 0.35), and the median loan was 58 million credits (interquartile range 37–110). The files contribute about 1,760 loan-years of observation, a median of 5.0 standard years per file.

By the close of the extract, 102 files (29.8%) had been restructured, 36 (10.5%) defaulted and collateralised, and 59 (17.3%) repaid as scheduled; 145 (42.4%) remained open. Censoring was heavy and uneven (Table 1). Only 6 of the 109 files opened in the earliest band were still open, against 94 of the 122 opened in the latest, and the crude proportion restructured fell across bands from 48.6% to 13.9% because late files had little time in which to be restructured. Allowing for follow-up, the Aalen–Johansen cumulative incidence at ten standard years was 40.0% (95% CI 33.3–47.0) for restructuring, 12.9% (9.2–17.5) for default and collateralisation, and 27.4% (20.7–34.7) for repayment as scheduled. The remaining 19.7% had no recorded ending within ten standard years.

In the primary model (Table 2), single-lane dependence carried the clearest association with restructuring: the hazard was 1.71 times that of files without such dependence (95% CI 1.14–2.56; p = .010). Each 0.10 rise in the DSBR was associated with a hazard 1.26 times higher (1.00–1.58; p = .049). The tribute indicator pointed the same way but was not distinguishable from no effect (1.25; 0.77–2.02; p = .37). Hazard ratios above 1 denote a higher rate of restructuring. Files with both features had a hazard 2.13 times that of files with neither (1.09–4.18; p = .028), an estimate that assumes the two effects multiply and was not tested against an interaction. Bootstrap standard errors were at most 8% larger than naive model-based errors, so clustering changed intervals only modestly.

None of the lender-side variables was clearly related to restructuring. Syndicated files had a hazard 0.93 times that of files where the Clan was sole lender (0.61–1.43), and each 0.25 increase in collateral coverage was associated with a ratio of 0.96 (0.83–1.12). A doubling of loan size gave 0.87 (0.75–1.01; p = .070), and sector governments had a hazard 1.44 times that of planetary governments (0.96–2.18; p = .079). Neither interval excludes 1, and we do not regard either as established. The opening cohort was not associated with the hazard (0.82 per five standard years later; 0.61–1.11), in line with the fall across bands in Table 1 arising from censoring.

The DSBR did not act evenly over time. In the first four standard years its hazard ratio was 0.99 per 0.10 (95% CI 0.73–1.36; 47 restructurings). After four years it was 1.56 (1.20–2.03; 55 restructurings). The ratio of the later to the earlier hazard ratio was 1.57 (1.09–2.26 from the joint bootstrap; p = .016). The pooled figure of 1.26 lies between the two and should be read as an average over unequal periods, not as a constant effect.

For default and collateralisation, with only 36 events, every interval included 1, and we make no claim that the predictors of the two endings differ or agree; the DSBR estimate, for one, was 1.19 (0.84–1.68).

Sensitivity analyses are gathered in Table 3. Treating all unresolved files as never restructured gave 28.9% (99 of 342) as the lowest possible ten-year proportion, and treating them all as restructured gave 68.1% (233 of 342), bounds too wide to constrain the Aalen–Johansen figure of 40.0%. Single-lane dependence kept a hazard ratio between 1.69 and 2.14 in the primary model and the four Cox-based sensitivity analyses, with intervals excluding 1 in each; the logistic odds ratio, on 208 resolved files, was 1.63 (0.89–2.99). The DSBR estimate moved more: it was 1.17 (0.94–1.46) with follow-up truncated at six years, 1.27 (0.86–1.88) in the 130 files with ten years of potential follow-up, and 1.24 (1.03–1.49) for the composite outcome. The logistic model rested on 208 files, of which 99 were restructured within ten standard years; that proportion (47.6%) is not an estimate of ten-year risk, because unresolved files were excluded. Recomputing the DSBR net of tribute gave 1.24 (0.99–1.55) and moved the tribute ratio to 1.12 (0.71–1.79).

Coding agreement was high. The two analysts disagreed on 6 of 120 files for single-lane dependence (kappa 0.90, 95% CI 0.82–0.98) and on 6 for tribute obligation (kappa 0.85, 0.74–0.97; 26 of the 120 files coded positive). The two analysts disagreed on 3 files for outcome state (kappa 0.94, 0.87–1.00, the upper limit truncated at 1). The intraclass correlation between the two abstractions of the DSBR was 0.98 (95% CI 0.96–0.99).

4. Discussion

Among the three candidate predictors, dependence on a single trade lane or export is the one on which we place most weight. Its hazard ratio was 1.71 in the primary model and stayed between 1.69 and 2.14 in the four Cox-based sensitivity analyses, with intervals excluding 1 in each; the logistic odds ratio, on 208 resolved files, was 1.63 (0.89–2.99). A borrower whose export earnings run through one lane has no cushion if that lane is disrupted, but the register cannot show that mechanism; the association is with the annex coding at opening, not with any event that followed.

The burden ratio is more difficult to summarise. The pooled hazard ratio of 1.26 has a lower limit at 1.00 and a p-value just under .05, which is weak support, and the split at four standard years shows why a single number misleads. Burden appears to matter little in the early years of a loan and more later. The six-year analysis (1.17; 0.94–1.46) mixes the two periods and is not independent confirmation. The change rests on one pre-specified split, and hazard ratios after four years are computed among files that survived, so depletion of the most fragile files may itself create an apparent change over time. We did not test other shapes of time dependence. What the data support is narrow: among files still open after four standard years, a higher opening burden is associated with a higher rate of subsequent restructuring.

For tribute, the point estimate was above 1, but its interval spans a modest reduction as well as a doubling of the hazard, so the data show neither that tribute leaves lending unaffected nor that it destabilises it. When the DSBR was recomputed with revenue net of tribute, the tribute estimate moved from 1.25 to 1.12 (0.71–1.79) with unchanged uncertainty, which is compatible with, but does not show, an effect through revenue. Obligations to a larger power may be better treated as a charge on revenue than as a flag, a suggestion for future measures and not a finding.

Syndication, collateral coverage and loan size were not distinguishable from no relation to restructuring, and their intervals were wide. The lender's own structuring choices, as coded in the register, did not visibly protect a loan against renegotiation. The size effect (0.87 per doubling) and the sector-government effect (1.44) are best treated as hypotheses for a larger register. The cause-specific hazards say nothing about the ten-year proportion restructured, which also depends on default and repayment; with 36 defaults we could not estimate that.

For practice, the results support a modest use of annex data already gathered at opening: closer scheduled review of files with single-lane dependence. Review of high-burden files from about the fourth standard year is a hypothesis for testing, not a recommendation.

5. Limitations

Selection is the principal limitation. The register holds only loans that were granted, and we cannot see applicants who were declined or who never applied (Mendrik, 33 BBY). No file has a DSBR above 0.57, and we do not know whether applications above that level were refused or simply not made. Selection on approval could bias the estimates in either direction; attenuation through restricted range is one possibility, not the expected one, and neither its size nor its direction can be estimated from these data.

Censoring remains unresolved. About 42% of files were still open at the closing date, and the primary and cumulative-incidence analyses assume that censoring is unrelated to the risk of restructuring given the covariates. The sensitivity analyses probe that assumption but cannot verify it.

The measures are ours. The 60% threshold for single-lane dependence was set in advance and was not varied, and the tribute indicator is generic. Across the three pre-specified hypotheses a Bonferroni threshold is .017; single-lane dependence (p = .010, adjusted .03) meets it and the DSBR (p = .049, adjusted .15) does not. Across all eight coefficients no result meets a strict threshold, and the secondary coefficients are exploratory. The study concerns one lender's sovereign lending and says nothing about other lenders, or about loans opened after 25 BBY. The design is observational: predictors were measured at opening, so they cannot be consequences of restructuring, but unmeasured features of borrowers may drive both.