Global Equipment Finance Scorecard

Designing a scorecard that crosses borders

Three things break when a scorecard leaves its home market. Only one of them is usually noticed.

1. The scales are incompatible — and percentiles fix it

FICO runs 300–850. CIBIL 300–900. Equifax Australia 0–1200. Credit Bureau Singapore 1000–2000. SCHUFA is a percentage. Even within the United Kingdom the three bureaus report on 0–999, 0–1000 and 0–710. A raw number is meaningless without knowing which bureau in which country produced it.

The distributions here are illustrative anchor points. In production you replace them with your own observed through-the-door population per market, which also means the normalisation self-corrects as a market’s credit quality shifts.

2. PAYDEX travels. “The business layer” does not.

The first version of this page said the business layer is portable and the consumer layer is not. That is too generous by half, and worth correcting in public. What is portable is PAYDEX specifically: a 1–100 index of dollar-weighted payment promptness that means the same thing everywhere D&B or a Worldwide Network partner operates. Eighty means paying on terms in Ohio and in Osaka. The Failure Score and Composite Credit Appraisal travel the same way, because they are the same vendor’s model on the same data.

Step outside D&B and the portability evaporates — and it evaporates in a way that is easy to miss, because the numbering survives. PAYDEX, Creditsafe’s International Score, Experian’s Intelliscore Plus and the Global Database credit score are all 1–100 and all point the same direction. They are not the same variable. PAYDEX reports what the business did. The other three report what a model thinks the business will do. Rapid Ratings’ FHR reports the shape of the balance sheet and never sees a trade line at all. A well-capitalised buyer that stretches its suppliers to 75 days scores badly on the first and well on the last, and nothing in the number tells you which one you are holding.

Nor is 1–100 the only collision. D&B’s Delinquency Predictor runs 101–670; NACM’s report runs 101–662. Different vendors, different models, different horizons, near-identical ranges. Experian’s Intelliscore Plus V3 runs 300–850 — the consumer FICO range, on a business. Cortera’s CPR runs 100–900 and reads the trailing three months, against PAYDEX’s twenty-four, so the same construct still translates worse than the shared definition implies. In every one of these cases the number arrives without its units attached.

Two entries break the pattern in instructive ways. NACM’s National Trade Credit Report is not on a 1–100 scale at all — it runs 101 to 662, which is close enough to consumer-score territory to be misread as one, and its horizon is six months rather than twelve. It is also the only source here built from a members’ cooperative: creditors contribute their own accounts-receivable ageing, and the score is computed on what members happened to report. That makes its trade data unusually direct and its coverage unusually conditional. A business with no NACM file is not thin-file; it simply does not buy from NACM members. Absence tells you about the supplier mix, not the buyer — which is exactly the sort of missingness a scorecard must not treat as a risk signal.

Global Database breaks it the other way. It carries the widest footprint of anything on this list — 600M+ companies across 200+ countries, aggregated from 400+ official registries plus local credit providers — and prices that breadth in consistency. One scale spanning 200 countries is not one model. Where filed accounts and trade data are deep the score behaves like Creditsafe’s; where they are thin it leans on registry facts and industry priors. So the same 65 is a different statement in Germany and in Ghana. It is comparable within a market and only loosely across them, which is the reverse of the property you actually want in a cross-border scorecard and the reason the market tier attribute exists separately.

Hence the bureau selector above. Choosing a non-native bureau does three things: it converts your raw score to a percentile inside that bureau’s own distribution, reads that percentile off the PAYDEX distribution to reach a bin the model understands, and then tells you what the conversion cost. Those costs are simulated on 400,000 paired draws under a Gaussian copula, with the rank correlation to PAYDEX set explicitly per bureau — 0.82 for another payment index, 0.38 for a statements-only score. The results are worth stating plainly:

Bureau	Scale	Construct	Horizon	ρ	Exact bin	Mean pts error
D&B PAYDEX	1–100	Payment behaviour	Trailing 24 months of payment experience	1.00	100%	0.0
Cortera Payment Rating (CPR)	100–900	Payment behaviour	Trailing 3 months	0.76	47%	7.8
D&B Delinquency Predictor	101–670	Failure forecast	12-month forecast	0.72	44%	8.4
NACM National Trade Credit Report	101–662	Failure forecast	6-month forecast	0.66	41%	9.3
Experian Intelliscore Plus	1–100 (V3: 300–850)	Failure forecast	12-month forecast	0.64	40%	9.6
Creditsafe International Score	1–100	Failure forecast	12-month forecast	0.58	37%	10.4
Global Database credit score	1–100	Failure forecast	12-month forecast	0.55	36%	10.8
CRIF / Cerved risk class	1–10 (1 = best)	Failure forecast	12-month forecast	0.52	35%	11.1
Rapid Ratings FHR	0–100	Accounts only	Latest filed accounts	0.38	31%	12.7

Three things stand out. First, even the closest substitute lands in the same bin only half the time — and that is a bureau measuring the same construct on a different panel. Second, the ranking follows the data source more than the vendor’s reputation: NACM, a members’ cooperative, translates better (ρ 0.66) than either of the large commercial forecasts below it, because trade lines are what PAYDEX is made of. Third, the measured bias is within ±0.06 points everywhere, so the translation is not systematically generous or harsh. The cost of a substitute bureau is variance, not drift. That is why the deal score is shown as a band when the source is not PAYDEX: the point estimate is honest as a centre and dishonest as a reading.

The practical consequence for a global book is that a cutoff calibrated on PAYDEX cannot be reused unchanged on a Creditsafe file. At a 600 cutoff, a business-layer error of ±10 points moves a material slice of the population across the line in both directions. Either recalibrate per bureau on local outcomes, or widen the refer band by roughly the mean points error and let a human read the file. Which markets actually leave you the choice is the subject of the global credit data coverage explorer — 50 of 103 markets there carry a single bureau, and in those you take the construct you are given.

3. Availability varies — so missing has to be a value

Some markets have no meaningful consumer bureau. Some businesses never file financials, so there is no Composite Credit Appraisal. Some jurisdictions weaken personal guarantees to the point of irrelevance. A model that requires a complete record cannot be deployed globally.

Every attribute therefore carries an explicit missing bin with its own empirically estimated weight of evidence — no imputation. “No D&B file in market” is scored on the observed default rate of deals with no D&B file, which in this sample runs 5.97% against 3.54% where a file exists. The scorecard degrades gracefully instead of failing.

The same mechanism does the segmentation work. Small-ticket deals never collect financial statements, so leverage and DSCR simply sit in their missing bins. One model serves all three ticket segments — validated separately on each: Gini 0.62 small, 0.68 mid, 0.67 large.

The finding I did not expect

Discrimination is not uniform across markets. The same scorecard achieves a Gini of 0.70 in the United States and 0.49 in Nigeria. That is not a modelling failure — it is a direct consequence of data depth. US deals arrive with a D&B file 93% of the time and a bureau score 89% of the time. Nigerian deals arrive with 36% and 41%.

The practical implication is uncomfortable and worth stating plainly: a global scorecard is not equally good everywhere, and pretending otherwise is how you lose money in frontier markets. Thin-data markets need a lower reliance on score, a heavier reliance on collateral and structure, and ideally a locally-fitted model once volume permits. The market risk tier attribute captures some of this, but it prices the environment, not the model’s own blindness in that environment.

What is deliberately excluded

Recovery. Repossession timelines, legal enforceability and secondary-market depth differ enormously across these fourteen markets, but those drive loss given default, not probability of default, and mixing them into a PD scorecard is a common and expensive error. Equipment resale depth appears here only as an origination-quality signal; the recovery half belongs in a separate LGD model and in the pricing, which is where the RAROC framework picks it up.

What would change in production

Per-market population distributions replacing the illustrative anchors. Segment-specific models once each segment carries enough defaults to support one. Reject inference. Fair-lending and data-protection review per jurisdiction — GDPR, India’s DPDP Act and Brazil’s LGPD all constrain what may enter a model and what must be explainable. And a governance decision most groups get wrong: who owns the cut-off. A single global cut-off applied to markets with different data depth quietly exports credit policy from headquarters to whichever market has the thinnest files.

Educational model built on synthetic data across illustrative market distributions. Not a production scorecard, not a lending recommendation, and not fit for credit decisioning. PAYDEX and D&B are trademarks of Dun & Bradstreet; FICO of Fair Isaac; other bureau marks belong to their owners. Runs entirely in your browser — nothing is stored or transmitted.