bureau-to-workout: simulated credit-bureau panels and debt workouts with ground truth

Every file in this dataset is synthetic. It comes from the two simulators of the paper, not from the National Credit Bureau (NCB) or Sukhumvit Asset Management (SAM). The simulators' parameters make the mechanisms visible at plausible magnitudes; nothing here is an estimate for Thailand. What the data offer that real data cannot is the ground truth: the outcomes of rejected applicants, the arrears a moratorium hides, and every borrower's response to every workout offer.

Data for the preprint From Bureau to Workout: Credit Scoring and Causal Debt Resolution under Policy-Shaped Labels by Thanatchon Chongmankhong (read online · PDF). The paper studies three ways the credit system writes its own labels, selection at approval, masking by deferral and treatment in the workout, and proposes an estimator for each. Code: https://github.com/thanatchon36/bureau-to-workout; archived on Zenodo: 10.5281/zenodo.24102251. Models trained on these data: thanatchon36/bureau-to-workout.

Independent personal research. Does not represent the views of any employer or institution.

from datasets import load_dataset

scoring = load_dataset("thanatchon36/bureau-to-workout", "scoring_e1")          # train / test, out of time
workout = load_dataset("thanatchon36/bureau-to-workout", "workout_cases", split="train")

Files

All draws are the paper's first seed, as make experiments makes them: the bureau without a moratorium with seed 7, the bureau with a moratorium with seed 1007, the workouts with seed 107. Settings are in configs.json.

path	content
scoring/e1_dev.parquet, scoring/e1_oot.parquet	the E1 modelling tables from the bureau without a moratorium: one row per scorable borrower (active, not yet 90+ DPD) at observation months 12–24 (train, 199,969 rows) and 36–45 (test, after the shock; 136,602 rows). 16 bureau attributes, y_obs (reported 90+ DPD within 12 months), y_true (the same without deferral), group columns and thin_file
scoring/e2_train.parquet, scoring/e2_test.parquet	the E2 reject-inference tables: every application to lender 0 in months 6–47 (9,531 train and 9,489 test rows, split by borrower), with the 15 bureau attributes the lender saw the month before, approved, observable (the bureau sees repayment on another account within 12 months, so a rejected applicant's outcome is still visible) and y_true, the applicant's 90+ DPD outcome had the loan been granted, for everyone. A lender sees y_true only where approved; the bureau also where observable
scoring/e3_dev.parquet, scoring/e3_oot.parquet	the E3 tables from the bureau with a moratorium: training through it (months 15–24, 160,195 rows) and testing after it (months 39–45, 108,817 rows). Adds y_delayed (windows touching a deferral extended to six months after exit) and mor_in_window
bureau/panel_{clean,moratorium}.parquet	the account-month panels, long format: 50,000 borrowers × 60 months. s_obs reported delinquency state (0 current, 1 30 DPD, 2 60 DPD, 3 90+ DPD, absorbing), s_true the state without deferral, mor deferred this month, util, n_acc, n_lenders, limit (THB), inq, new_acc
bureau/borrowers_*.parquet	static attributes: age at month 0, gender, region, occupation, new_to_credit_at_start; ability and soft are latent (lenders' soft information), for evaluation only
bureau/applications_*.parquet	every credit application: borrower, month, lender (0 is the lender studied in E2, 5 the lenient lender), approved (86,740 and 90,384 rows)
workout/cases.parquet	50,000 NPL cases: 13 covariates from the AMC file and the bureau, balance, the logged offer (0 standard, 1 term extension, 2 write-down, 3 both), realised recovery (present value ÷ balance), officer, and the truth for every offer k: propensity_k, value_k (expected recovery), p_complete_k, spill_k (new 90+ DPD elsewhere within 24 months); latent capacity and willingness; split as in the paper's R1–R4
results/	every result file of the paper, five seeds (make experiments): the CSVs behind each table and figure
configs.json, versions.json	simulator settings, and the library versions used to build this upload
SHA256SUMS	checksums of every file above

Tasks these data support

• Reject inference / selective labels. In scoring_e2_reject_inference, train on approved rows only (what a lender has), on approved | observable rows (what the bureau adds), or on everything (the oracle), and score the test split against y_true, including rejected applicants.
• Label masking. In scoring_e3_moratorium, compare models trained on y_obs, y_delayed and y_true, and test after the moratorium.
• Heterogeneous treatment effects and policy learning. In workout_cases, estimate offer effects from the logged offers, then check them against value_k; learn a write-down-budgeted policy and evaluate it exactly.
• Off-policy evaluation. Estimate a policy's value from logged data and compare with its true value, using propensity_k (true) or your own estimate.

Licence

CC BY 4.0 (see LICENSE). The code that generates the data is MIT-licensed.

Citation

@misc{chongmankhong2026bureau,
  author = {Chongmankhong, Thanatchon},
  title  = {From Bureau to Workout: Credit Scoring and Causal Debt Resolution under Policy-Shaped Labels},
  year   = {2026},
  doi    = {10.5281/zenodo.24102251},
  url    = {https://thanatchon36.github.io/bureau-to-workout/paper/},
  note   = {Preprint; all data are synthetic}
}