hindsight-or-foresight: coded failure cases, alternative-data scores and hindsight-bias results

Verify before citing. data/cases_coded.csv was coded from the author's knowledge of the public record without news-database access: dates are at monthly resolution and each case carries a confidence flag (High / Medium / Low). data/taxonomy_scores.csv is a structured judgement, not a measurement. Check the cases, especially the Thai cases flagged Low, against SET and SEC Thailand news, court records and company filings before relying on them.

Data for the preprint Hindsight or Foresight? Alternative Data, Early Warning, and the Measurement of Corporate-Group Failure: Global Evidence and Implications for Thailand by Thanatchon Chongmankhong (read online · PDF). The paper asks how much of "the signs were there" survives once hindsight is removed. It scores eleven families of alternative data, codes 28 failures and turnarounds by the first public signal in each family, measures hindsight bias on 146,045 U.S. firm-years with SEC enforcement labels, and isolates five bias channels with Monte Carlo simulations. Code: https://github.com/thanatchon36/hindsight-or-foresight; archived on Zenodo: 10.5281/zenodo.24216338. Model specifications: thanatchon36/hindsight-or-foresight (model).

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

Files

path	content
data/cases_coded.csv	The case corpus. 28 firms and business groups (24 failures, 4 turnarounds; 22 global, 6 Thai): country, region, sector, the month of the first public signal (t_signal), of revelation (t_reveal) and of the terminal event (t_terminal), each with a note; a confidence flag; and nine columns, one per signal family, coded 0 (no documented public signal before revelation), 1 (a public signal less than twelve months before) or 2 (twelve or more months before). For turnarounds the signal is the first recovery signal, t_reveal the distress date and t_terminal confirmed recovery
data/taxonomy_scores.csv	The taxonomy. Eleven alternative-data families (F1 to F11) scored 1 (weak) to 5 (strong) on lead time, specificity, global access, Thai access, cost, legal safety under Thailand's PDPA and point-in-time integrity
results/e1_ladder.json	E1, the leakage ladder: ROC AUC, average precision, precision and recall in the top 1% and their bootstrap 95% intervals for gradient boosting and logistic regression under four evaluation designs, test years 2003 to 2008
results/e2_sampling.json	E2, sampling design: precision of the honest model's threshold in case–control samples with 1 to 50 controls per case and in the full population
results/e3_event_time.csv, e3_meta.json	E3, event time: mean within-year percentile of the honest risk score (gradient boosting, 2-year gap, serial-case exclusion) by year relative to the first misstated year, for sanctioned firms and size-matched controls (163 cases, 6,753 firm-years)
results/e4_censoring.csv	E4, label right-censoring: firm-years, sanctioned firm-years and their rate by fiscal year
results/e5_groups.csv	E5, feature groups: drop in honest ROC AUC when each group of accounting items is permuted, by test year
results/s1_forking.csv	S1, forking paths: in-sample hit rate on known failures of the best rule found among K noise signals, against its out-of-sample hit and false-alarm rates
results/s2_alignment.json	S2, alignment artefact: paths aligned on the failure date, for failing firms, placebo dates and controls (mean and 10th / 90th percentiles by event time)
results/s3_timeliness.csv, s3_meta.json	S3, timeliness against noise: share of failures warned at each lead time and false-alarm rate, for audited accounts with a six-month reporting lag and alternative data at noise sd 0.05, 0.15 and 0.30
results/s4_ppv.csv	S4, base rates: positive predictive value over a grid of sensitivity, specificity and base rate
results/real_log.txt	the run log of the real-data experiments
SHA256SUMS	checksums of every file above

E1 to E5 use the U.S. panel; S1 to S4 are simulations with fixed seeds. The paper's figures and tables are built from these files by code/figures.py and code/tables.py.

What is not here

The firm-year panel itself, and the firm-level score files built from it (honest_predictions_2003_2008.csv, rolling_predictions_1997_2014.csv, keyed by Compustat gvkey), are not redistributed on the Hub. The panel is the public release of Bao, Ke, Li, Yu and Zhang (2020), "Detecting Accounting Fraud in Publicly Traded U.S. Firms Using a Machine Learning Approach", Journal of Accounting Research 58(1), 199–235 (doi:10.1111/1475-679X.12292), distributed at https://github.com/JarFraud/FraudDetection without a stated licence. Get it from there, or with bash code/get_data.sh in the code repository, which also holds the firm-level score files.

Licence

The case corpus, the taxonomy scores and all results are released under CC BY 4.0 (see LICENSE). The results E1 to E5 are aggregates computed from the Bao et al. (2020) panel; cite their paper when you use them.

Citation

@misc{chongmankhong2026hindsight,
  author       = {Chongmankhong, Thanatchon},
  title        = {Hindsight or Foresight? Alternative Data, Early Warning, and the Measurement of Corporate-Group Failure: Global Evidence and Implications for Thailand},
  year         = {2026},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.24216338},
  url          = {https://thanatchon36.github.io/hindsight-or-foresight/paper/}
}