thanatchon36/finality-research

finality: simulated data for extreme-value uncertainty and contagion in banking-sector cyber risk

Every number here is simulated from illustrative parameters. Nothing in this dataset is an observed cyber loss, a real bank or a real exposure. The data exist to audit statistical methods against a known truth and to reproduce the paper; recalibrate with real incident data before citing any number.

Data for the preprint False Confidence in the Tail: Finality, Extreme-Value Uncertainty, and Contagion in Banking-Sector Cyber Risk by Thanatchon Chongmankhong (read online · PDF). The paper asks how confidently one can state the return period of a catastrophic banking-sector cyber loss from a few decades of data. A loss process with a planted Pareto tail gives a known truth; 36,000 interval estimates for the 1-in-100 and 1-in-1,000-year loss are scored against it. A rational-choice attacker model and a linear DebtRank interbank network complete the framework. Code: https://github.com/thanatchon36/finality-research; archived on Zenodo: 10.5281/zenodo.24050729.

Independent personal research. Defensive, actuarial and policy research: the models describe attacker incentives and aggregate losses only, and the dataset contains no attack technique.

Files

path	content
coverage/intervals.csv	The coverage study, record by record (36,000 rows): for each experiment (A, B), record length, simulated record (1 to 1,000), return period (100, 1,000 years) and method (delta, profile, bootstrap): the number of threshold exceedances, the point estimate, the nominal 95% interval, the true return level, and whether the interval covered the truth or lay wholly below it (truth_above) or above it (truth_below). Records whose fit failed (fewer than eight exceedances) are kept and marked fit failed
coverage/summary.csv	the 36 cells of the study: coverage with its Monte Carlo standard error, miss rates on each side, median upper/lower ratio, share of unbounded intervals, median point estimate over truth, failed records
coverage/designs.json	the two designs, their record lengths, threshold rules and true return levels
loss/annual_losses_200k.csv	200,000 simulated years of the banking-sector loss process: incidents per year and the annual aggregate loss (USD)
loss/truth_return_levels.csv	the true annual return-level curve from 2,000,000 simulated years, return periods 1.25 to 100,000 years
loss/record_1000y_incidents.csv, loss/record_100y_incidents.csv	the two single records analysed in Section 5.3 of the paper: every incident loss with its year
loss/parameters.json, loss/risk_metrics.json, loss/single_record_fits.json	the loss-process parameters (planted tail index 0.625), VaR and expected shortfall of the 200,000-year sample, and the peaks-over-threshold fits to the two records
attacker/venues.csv	the four attack venues: breach probability, value at risk, recovery probability (one minus finality), laundering haircut, attribution probability, sanction, cost, retained value, expected payoff and break-even recovery
attacker/log_gap_bank_to_bridge.csv	the exact decomposition of the log payoff gap between a bank core and a cross-chain bridge (finality: 67%)
attacker/recovery_curves.csv	each venue's expected payoff as its recovery probability varies from 0 to 0.95
contagion/banks.csv	the 28-bank core-periphery network (seed 7): tier, total assets, equity, interbank intensity, shared-vendor client flag and contagion multiplier when the bank alone loses 20% of equity
contagion/exposures.csv	the 92 interbank exposures: weight is the share of the creditor's interbank assets lent to the debtor (rows sum to one)
contagion/phase_sweep.csv	the contagion multiplier against the exposure scale for a hit on the largest bank and on a shared vendor; the critical scale is 0.868
contagion/cascades.csv	the three scenarios of the paper round by round: every bank's distress (0 to 1) after each propagation step
raw/	the study's own outputs: results.json, results.md and the per-record interval cache in raw/cache/
SHA256SUMS	checksums of every file above

How the data were generated

scripts/export_hf.py in the code repository regenerates every file from master seed 20261108 exactly as run_study.py draws it, and refuses to write unless each table reproduces the stored results to nine significant digits.

• Loss process. Incidents arrive as a Poisson process (4 a year). Severities are log-normal (median $30M, sigma 1.4) with probability 0.95 and Pareto (scale $200M, alpha 1.6, so xi = 0.625) with probability 0.05; an incident above $1B triggers a cascade with probability 0.3, multiplying its loss by one plus an exponential with mean 2.
• Experiment A draws exceedances exactly from a generalized Pareto distribution (threshold $400M, sigma $250M, xi 0.625, 0.5 exceedances a year) for 30, 100 and 300-year records: the model is exactly right.
• Experiment B simulates the full loss process for 30, 100 and 1,000-year records and fits a GPD above the 85th, 95th and 98th percentiles of incident losses; return levels are corrected from incident level to annual aggregate.
• Intervals: delta method on the log scale, profile likelihood, and parametric bootstrap with B = 199.
• Network: a 5-core, 23-periphery network from the same random number generator as the web instrument; linear DebtRank with a 20% equity shock.

Licence

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

Citation

@misc{chongmankhong2026false,
  author = {Chongmankhong, Thanatchon},
  title  = {False Confidence in the Tail: Finality, Extreme-Value Uncertainty,
            and Contagion in Banking-Sector Cyber Risk},
  year   = {2026},
  doi    = {10.5281/zenodo.24050729},
  url    = {https://thanatchon36.github.io/finality-research/paper/},
  note   = {Preprint; all parameters are illustrative}
}