shadow-dollars: stablecoin rails, sanctions and detection data for Thailand

Two kinds of data. raw/ and processed/ are real public data (on-chain metrics, exchange rates, inflation, sanctions lists) and results derived from them. simulation/ is simulated: Monte Carlo results from an agent-based exchange simulator with planted mule accounts. It validates detection methods against known ground truth and is not evidence about real markets. No account-level Thai exchange data are included.

Data for the preprint Shadow Dollars: Stablecoin Rails, Scam Finance and the Limits of Detection in Thailand by Thanatchon Chongmankhong (read online · PDF). The paper asks whether Thai stablecoin behaviour is anomalous, whether USDT is a plausible channel for laundering transnational fraud proceeds, and how either claim can be tested credibly. It finds that USDT on Tron is a retail dollar rail distinct from USDT on Ethereum, that chain-wide data cannot evaluate national enforcement, and it validates four account-level detectors in simulation. Code: https://github.com/thanatchon36/shadow-dollars; archived on Zenodo: 10.5281/zenodo.24050717.

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

Files

path	content
raw/cm_*.csv	real. Coin Metrics Community Data, one file per asset or asset-on-chain, daily: supply (SplyCur), addresses with a balance (AdrBalCnt), active addresses (AdrActCnt), transaction and transfer counts (TxCnt, TxTfrCnt), price and market cap. Stablecoins: USDT in aggregate and on Tron (usdt_trx, from 2019), Ethereum (usdt_eth), Omni and Avalanche C-chain; USDC in aggregate, on Ethereum and on Tron; DAI, BUSD, PYUSD, FDUSD, USDe. Reference assets: BTC, ETH, TRX. Through 23–24 May 2026
raw/fx_daily.csv	real. Federal Reserve H.10 daily noon buying rates in long format (Date, Country, Exchange rate), 1971 to 30 October 2026, via datasets/exchange-rates. Most series are local currency per US dollar; the euro, pound, Australian and New Zealand dollars are US dollars per unit
raw/wb_cpi.csv	real. World Bank FP.CPI.TOTL.ZG, annual consumer-price inflation (%), via datasets/inflation. That mirror swaps the CPI and GDP-deflator file labels; this is the CPI file (published there as inflation-gdp.csv), checked against known prints (Thailand 2022 = 6.1%, United States 2022 = 8.0%)
raw/ofac/*.txt	real. OFAC SDN digital-currency addresses, one file per currency tag (17 tags, 1,073 addresses), via 0xB10C/ofac-sanctioned-digital-currency-addresses
raw/MANIFEST.json	URL, retrieval time (8 November 2026), size and SHA-256 of every raw file as downloaded. It also lists two World Bank files not included here: a duplicate of wb_cpi.csv and the GDP deflator
processed/onchain_snapshot.csv	the rails on 23 May 2026 (paper Table 1): supply (US$ bn), holders (millions), active addresses and transfers (thousands a day), 1-year % changes (*_g) and mean balance per holder (US$), for USDT on Tron and Ethereum and USDC on Ethereum and Tron
processed/cpi_2019_2024.csv	mean, min, max and number of years of annual CPI inflation, 2019–2024 where available (Thailand to 2023), for 23 comparators
processed/fx_vol_2021_2026.csv	annualised volatility (%) of daily log changes against the US dollar, January 2021 to October 2026, 21 currencies
processed/ofac_by_format.csv	sanctioned addresses by currency tag and address format (Bitcoin, EVM 0x…, Tron T…, other)
processed/event_study_results.csv	three Thai shocks: 30-day cumulative abnormal log USDT-on-Tron transfers (car30), the nominal 95% band (lo, hi) and the p-value calibrated on the placebo dates
processed/event_placebos.csv	the same event study at 149 placebo dates (every 7 days, away from real events); outside marks dates where the nominal band rejects (58%)
processed/facts_real.json	every public-data number quoted in the paper
simulation/mc_ledger.csv	simulated. One row per simulated exchange-year: mule share (share: 0 for the null, 200 draws; 0.25–4%, 50 draws each), seed, number of mules, the difference-in-bunching estimate (b_usdt, b_other, delta_b, se_delta, p_bunch, reject_bunch), and AUC and precision at k (auc_*, p@k_*) for seven detectors: dwell time only, undirected two-hop exposure, directed three-hop exposure, transparent composite scores and Isolation Forests (behaviour only, or all signals)
simulation/mc_scm.csv	simulated. Synthetic control on simulated country panels: true effect (log points; 0 to 0.3, 200 panels each), seed, estimated ATT, placebo p-value, rejection at 5%, pre-treatment RMSPE
simulation/facts_sim.json	every simulation number quoted in the paper (size, power, precision at k, AUC)
SHA256SUMS	checksums of every file above

Reproducing

git clone https://github.com/thanatchon36/shadow-dollars && cd shadow-dollars
make setup && make all    # data -> figures -> Monte Carlo (~10 min) -> simulation -> tests -> paper -> web

make data downloads only missing files, so placing raw/ from this dataset in data/raw/ reproduces the paper's vintage exactly; downloading afresh gives a later vintage, since some sources update daily. Simulation seeds are fixed, so simulation/ reproduces exactly.

Licence and attribution

LICENSE gives the terms file by file. In short: the Coin Metrics files and everything derived from them are CC BY-NC 4.0 (non-commercial, attribution to Coin Metrics); the World Bank file is CC BY 4.0; the Federal Reserve and OFAC data are U.S. government works; the simulation results and the author's other derived tables are CC BY 4.0.

No individual is accused of wrongdoing. The sanctions addresses are published by the U.S. Treasury; no address of a private person is de-anonymised.

Citation

@misc{chongmankhong2026shadowdollars,
  author = {Chongmankhong, Thanatchon},
  title  = {Shadow Dollars: Stablecoin Rails, Scam Finance and the Limits of Detection in Thailand},
  year   = {2026},
  doi    = {10.5281/zenodo.24050717},
  url    = {https://thanatchon36.github.io/shadow-dollars/paper/},
  note   = {Preprint}
}