mkbadeniyi/fraud-detection-benchmark

Fraud Detection Benchmark

Five fraud datasets. One schema. Split the way the real world happens.

Most fraud benchmarks shuffle time and leak the answer. This one doesn't. Every dataset is cleaned, checked for leaks, and cut into train, validation and test strictly by time, so the score you see is the score you'd get in production.

 

Inside

	Dataset	Fraud type	Kind	Rows	Fraud
🇺🇬	Xente	Mobile money	Real	95,662	193
💳	IEEE-CIS	Card payments	Real	590,540	20,663
📱	PaySim	Mobile money	Synthetic	6,362,620	8,213
📱	MoMTSim	Mobile money	Synthetic	4,225,958	2,233,118
🏦	BAF	Bank account opening	Synthetic	1,000,000	11,029

12.3 million rows. Two real. Three synthetic. One file each.

 

Get started

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/mkbadeniyi/fraud-detection-benchmark/"
    "fraud_datasets/xente_mobile_money_uganda_real.parquet"
)
train, test = df[df.split == "train"], df[df.split == "test"]

Or with 🤗 Datasets:

from datasets import load_dataset

ds = load_dataset("mkbadeniyi/fraud-detection-benchmark", "xente", split="full")

 

One schema

Every file opens with the same six columns. Each dataset's own features follow.

Column	Type	What it means
dataset	category	Where the row comes from
split	category	train, validation or test, cut by time 70 / 15 / 15
ts	time	When it happened, in the source's own unit
label	int8	1 is fraud
user_id	string	The account, where the source has one
seen_user	boolean	This user appeared in an earlier split

Rows are already in time order.

 

Built with care

• No leaks. Columns that give the answer away are gone: PaySim's and MoMTSim's balance columns, and PaySim's isFlaggedFraud.
• Time-true splits. Train comes first, then validation, then test. Never shuffled.
• Returning users, flagged. Users seen before keep their rows and get seen_user, so you can score new users on their own.
• Verified. Row counts, fraud counts and time order are checked against the source on every build.

 

Good to know

• Fraud rises over time in Xente, PaySim and BAF. Random splits will flatter you.
• Xente is small. About 100 fraud rows in train, all above 200,000 UGX. Amount alone ranks it almost perfectly.
• MoMTSim is easy on purpose. Every fraud is a TRANSFER. Use it to test pipelines, not to claim results.
• A near-perfect score is a leak until you prove it isn't.

 

Tech specs

File	Rows	Columns	Train / Val / Test fraud rate	ts unit	Size
xente_mobile_money_uganda_real	95,662	19	0.16% / 0.27% / 0.35%	UTC datetime	2.3 MB
ieee_cis_card_payments_real	590,540	438	3.52% / 3.43% / 3.48%	seconds from reference	61.7 MB
paysim_mobile_money_synthetic	6,362,620	9	0.08% / 0.06% / 0.44%	simulation hour	112.8 MB
momtsim_mobile_money_synthetic	4,225,958	9	53.8% / 51.0% / 50.3%	simulation step	44.0 MB
baf_bank_account_opening_synthetic	1,000,000	36	1.03% / 1.34% / 1.47%	month	66.0 MB

Parquet, Zstandard compressed.