fraud-synth-bench: synthetic transaction data

Every dataset behind the paper Classic and Modern Fraud Detectors under Controlled Imbalance, Drift, Ring Fraud and Label Delay: A Reproducible Study on Synthetic Data (Chongmankhong, 2026): 45 generated datasets (9 experimental conditions × 5 seeds), about 200,000 transactions each, plus the per-run results of the experiments.

📄 Paper (web) · PDF · 💻 Code · 🔖 DOI 10.5281/zenodo.23343483

This is synthetic data. No real transactions, customers or institutions were used to build, fit or calibrate the generator, and it contains no PII. Distributions are stylised from qualitative descriptions in the literature. Findings on this data are evidence about mechanisms, not about any real payment system. Absolute AUPRC values should not be compared with published numbers on real data.

What's here

Condition (data/<condition>/seed{0..4}/)	Preset	Fraud rate	Manipulated	Used by
small_fr0.001	small	0.1%	extreme imbalance	RQ1
small_fr0.005	small	0.5%	baseline	RQ1, RQ5
small_fr0.01	small	1%	imbalance	RQ1
small_long_drift	small_long	0.5%	fraud tactics switch at day 54	RQ2
small_ring0.0, small_ring0.25, small_ring0.5, small_ring0.75	small	0.5%	share of fraud from ring scenarios	RQ3
small_long	small_long	0.5%	baseline for label delay (the delay itself is in labels)	RQ4

small has 20,000 accounts over 30 days. small_long has 8,000 accounts over 90 days. The total is 8.9M transactions, stored in about 190 MB. small_fr0.005 and small_long are also what make data produces.

Other files:

• index.parquet: one row per dataset, with its path, sizes and full content_hash.
• results/runs/rq{1..4}.jsonl: one line per model × condition × seed, with metrics and the dataset_hash of the data it ran on. RQ5 is computed from the RQ1 runs.
• results/summary.json, results/summary/: the aggregates that fill in the paper's numbers. results/skipped.json records the skipped model (TabPFN).
• docs/DATA_CARD.md: the full data card (simulation details, validation, limitations). docs/DATA_CARD_STATS.md: descriptive statistics.

Tables

Each data/<condition>/seed<n>/ directory holds:

File	Grain	Columns
transactions.parquet	one row per transaction	tx_id, timestamp, account_id, merchant_id (null for transfers), counterparty_id (null for card payments), device_id (null for card-present), amount, mcc, channel (pos / online / transfer), country
labels.parquet	one row per transaction	tx_id, is_fraud, observed_label, label_available_at, scenario, fraud_event_id
accounts.parquet	one row per account	account_id, opened_at, home_country, segment, account_role
merchants.parquet	one row per merchant	merchant_id, mcc, country, opened_at
edges.parquet	one row per (src, dst, type)	src_type, src_id, dst_type, dst_id, edge_type (pays / transfers / uses_device), first_seen, last_seen, n_tx
manifest.json	per dataset	generator version, preset, seed, all parameters, content hash, summary stats, library versions

IDs are only meaningful within one dataset. Don't pool seeds or conditions without keying on them.

Fraud comes from eight scenarios: card_testing, account_takeover, velocity_abuse, friendly_fraud, bust_out, and three ring scenarios, mule_network, merchant_collusion and synthetic_identity.

⚠️ Avoiding leakage

• labels.is_fraud, labels.scenario, labels.fraud_event_id and accounts.account_role reveal the answer. Use them for evaluation and analysis, never as model features.
• Train on observed_label only where label_available_at ≤ your training cutoff. Undetected fraud keeps observed_label = 0 forever.
• edges is aggregated over the whole timeline, so using it directly as a feature source leaks the future. Build time-respecting graph features from transactions with a cutoff.
• Split by time, not at random.

Usage

The viewer configs cover the two baseline conditions, with one split per seed:

from datasets import load_dataset

tx = load_dataset("thanatchon36/fraud-synth-bench", "small_transactions", split="seed0")
labels = load_dataset("thanatchon36/fraud-synth-bench", "small_labels", split="seed0")

Any condition or table can be read by path:

import polars as pl

base = "hf://datasets/thanatchon36/fraud-synth-bench/data/small_long_drift/seed0"
df = pl.read_parquet(f"{base}/transactions.parquet").join(
    pl.read_parquet(f"{base}/labels.parquet"), on="tx_id"
)

To use a whole directory with the project's loader, which restores the column types and exposes the content hash:

from huggingface_hub import snapshot_download
from fraudlab.synth import load  # https://github.com/thanatchon36/fraud-synth-bench

root = snapshot_download("thanatchon36/fraud-synth-bench", repo_type="dataset",
                         allow_patterns="data/small_ring0.5/seed0/*")
ds = load(f"{root}/data/small_ring0.5/seed0")

Reproducing it

The output is a pure function of the config. These are the commands that produced this upload (with --seed 0 to 4, generator version 0.1.0):

fraudlab generate --preset small --fraud-rate 0.005 --seed 0 --out data/small_fr0.005/seed0
fraudlab generate --preset small --fraud-rate 0.005 --ring-share 0.25 --seed 0 --out data/small_ring0.25/seed0
fraudlab generate --preset small_long --drift rq2 --seed 0 --out data/small_long_drift/seed0
fraudlab generate --preset small_long --fraud-rate 0.005 --seed 0 --out data/small_long/seed0

The content_hash covers column names, dtypes and values, not parquet writer metadata. All 45 datasets here were checked against the dataset_hash values recorded in results/runs/*.jsonl, and all match. The tiny (20k transactions) and full (2M) presets aren't in the paper and aren't uploaded. The commands above generate them.

Limitations

1. Not calibrated to real data. Real data has heavier tails, more merchant categories, and seasonality that the generator doesn't model.
2. Fraud is more mechanistic than in reality. Each scenario has a clean signature, and most are easier to detect than real fraud, except friendly_fraud, which is close to undetectable by design.
3. Few ring events. At small with 0.5% fraud, ring scenarios contribute about 4–8 events each, so per-scenario results have high variance.
4. No adaptive adversary. Drift is scheduled, not a reaction to a deployed model, and legitimate behaviour doesn't drift.
5. Single currency and abstract countries (C00–C11).

Not intended for estimating real-world fraud prevalence or losses, benchmarking vendor products, or training production models.

Citation

@software{chongmankhong2026fraudsynthbench,
  author = {Chongmankhong, Thanatchon},
  title  = {fraud-synth-bench: Reproducible fraud-analytics research on synthetic data},
  year   = {2026},
  doi    = {10.5281/zenodo.23343483},
  url    = {https://github.com/thanatchon36/fraud-synth-bench}
}

License

Apache-2.0, the same as the code.