Generate Credit Risk Synthetic Data in Python

Every bank capital model runs on three numbers: probability of default (PD), loss given default (LGD), and exposure at default (EAD), combined into expected loss as EL = PD × LGD × EAD. A credit_score column and a defaulted column with no statistical relationship between them is worse than useless for validating a risk model — realized defaults have to actually track the declared PD, by rating grade, for the data to mean anything. Misata generates a loan portfolio where they do, using the regulator's own published numbers rather than plausible-sounding placeholders.

import misata

schema = {
    "borrowers": {
        "__rows__": 3000,
        "borrower_id": {"type": "integer", "primary_key": True},
        # S&P's real credit-rating scale, not a fictional 1-10 score.
        "credit_rating": {"type": "string",
            "enum": ["AAA", "AA", "A", "BBB", "BB", "B", "CCC"],
            "weights": [0.02, 0.06, 0.17, 0.30, 0.22, 0.16, 0.07]},
    },
    "loans": {
        "__rows__": 4200,
        "loan_id": {"type": "integer", "primary_key": True},
        "borrower_id": {"type": "integer", "foreign_key": {"table": "borrowers", "column": "borrower_id"}},
        "seniority": {"type": "string", "enum": ["senior_unsecured", "subordinated"], "weights": [0.82, 0.18]},
        "drawn_amount": {"type": "float", "distribution": "lognormal", "mu": 13.5, "sigma": 1.1,
                          "min": 50_000, "max": 25_000_000, "decimals": 2},
    },
}
tables = misata.generate_from_schema(misata.from_dict_schema(schema, seed=11))
print(list(tables.keys()))   # ['borrowers', 'loans']

That is the minimal shape. The full example — PD by rating, LGD by seniority, EAD from drawn plus a credit-conversion-factor on undrawn commitment, and a real Bernoulli default draw at each borrower's own PD — is a working, runnable script: examples/credit_risk_portfolio.py in the repo. Run it directly:

python examples/credit_risk_portfolio.py

It prints every guarantee below, checked against the data it just generated:

borrowers: 3000  loans: 4200
portfolio EAD: $5,629,505,202   portfolio expected loss: $62,364,849   (1.108% of EAD)

  [OK] 'AAA' (S&P PD 0.00%): exactly zero realized defaults (0 of 68)
  [OK] 'AA' (S&P PD 0.00%): exactly zero realized defaults (0 of 281)
  [OK] 'A' (S&P PD 0.00%): exactly zero realized defaults (0 of 718)
  [OK] 'BBB' realized default rate 0.0016 vs S&P PD 0.0003
  [OK] 'BB' realized default rate 0.0033 vs S&P PD 0.0028
  [OK] 'B' realized default rate 0.0141 vs S&P PD 0.0161
  [OK] 'CCC' realized default rate 0.2803 vs S&P PD 0.2692
  [OK] 'senior_unsecured' LGD is exactly 40% on every loan
  [OK] 'subordinated' LGD is exactly 75% on every loan
  [OK] EAD reconciles to drawn + CCF x undrawn on every loan
  [OK] term loans have zero undrawn commitment
  [OK] expected_loss equals PD x LGD x EAD on every loan
  [OK] realized_loss is zero for every non-defaulted loan
  [OK] realized_loss equals LGD x EAD for every defaulted loan
  [OK] loans.borrower_id has zero orphans

ALL CHECKS PASSED

What each number is grounded in

PD, by credit rating. S&P Global Ratings' own published annual global corporate default rate, averaged across the six most recent studies (2019-2024) rather than any single volatile year. AAA, AA, and A each show exactly 0.00% in every one of those six years — a real fact about investment-grade defaults, not a placeholder. This example treats that as a guarantee to verify (exactly zero realized defaults in those three grades, checked against the actual generated rows) rather than smoothing it into something that looks more "generator-friendly."

LGD, by seniority. The Basel Foundation IRB supervisory value: 40% for senior unsecured claims on corporates, 75% for subordinated claims. This is not an estimate — it is the number a bank using the F-IRB approach is required to use (BIS/OSFI Capital Adequacy Requirements, Chapter 5, Internal Ratings-Based Approach), which makes it the least arguable figure in the whole example.

EAD, for a partially-drawn facility. EAD = drawn + CCF × undrawn, the Basel credit-conversion-factor formula. CCF is 20% for commitments of one year or less, 50% for longer commitments, and 0% for commitments that are unconditionally cancellable at any time — the same table used in the Basel standardized approach. A term loan is fully drawn by construction, so its undrawn commitment is exactly zero, not a stray random value the schema happened to assign.

Expected loss. EL = PD × LGD × EAD, recomputed independently in verify() from the raw columns, not just declared.

The connection. A borrower's credit_rating is not a decorative label sitting next to an independently-random defaulted flag. Whether a loan actually defaults is a real Bernoulli draw at that borrower's own S&P-published PD, so the realized default rate per rating grade is a measured property of the generated rows — checkable, not assumed.

Portfolio composition

The rating-grade mix (2% AAA, 6% AA, 17% A, 30% BBB, 22% BB, 16% B, 7% CCC) is declared to be realistic of the real corporate bond market's skew toward investment grade, not independently cited the way the PD table itself is — stated plainly, the same honesty split used for the designed severity-tier progression in the healthcare comorbidity example.

A realism bug found and fixed after this shipped

The first release of this example passed every guarantee above and still gave itself away: drawn_amount came out as 130028.33004434995, and origination_date as 2025-09-12 08:16:54. No loan tape on earth carries a dollar figure to eleven decimal places or an origination date to the nanosecond. The first was a one-line schema fix ("decimals": 2). The second was a real engine bug — "type": "date" was silently getting a time-of-day added by the same temporal-profile system that gives "type": "datetime" columns their realistic business-hour grids, in four separate code paths in misata/simulator.py. Fixed in 0.9.6.36, with a permanent regression test at the engine level (tests/test_date_type_has_no_time.py) so any schema declaring "type": "date", not just this one, stays a calendar day rather than a timestamp.

What this is not

This models unsecured commercial lending using the Foundation IRB's own supervisory parameters, not a full internal-ratings-based capital calculation. Real F-IRB models also incorporate maturity adjustments, correlation parameters, and a bank's own internal PD estimation rather than S&P's public corporate-bond default study — this example is grounded in real, checkable regulatory numbers to produce realistic test data, not a substitute for an actual capital adequacy filing. Retail/consumer lending (mortgages, credit cards held by individuals) uses different LGD conventions than the corporate F-IRB values used here; this example is scoped to commercial/corporate lending specifically.