Prompt
I'm building a fraud detection app and need transaction data with labels and enough history. What datasets are worth looking at?
Latest observation
Oct 16, 2026
OpenAI APIWeb search: off
For a fraud detection app, you’ll usually want datasets with:
- Transaction-level records
- A fraud label (
fraud/legit) - Enough temporal history to model behavior over time
- Ideally multiple entities: users, cards, merchants, devices, IPs, timestamps, amounts
Here are some of the most useful datasets to look at, grouped by type.
Commonly used benchmark datasets
1) Kaggle Credit Card Fraud Detection
- What it is: European card transactions with fraud labels
- Size: ~284k transactions
- Strengths: Very popular benchmark, labeled, easy to use
- Limitations: Highly anonymized features, little real-world context, no rich entity history
- Best for: Baseline models, imbalanced classification experiments
2) IEEE-CIS Fraud Detection
- What it is: E-commerce transaction + identity data
- Strengths: Richer features than the Kaggle credit card set, includes transaction and identity tables
- Limitations: Still anonymized, somewhat messy, not ideal for direct business interpretation
- Best for: Feature engineering, entity linking, heterogeneous tabular models
3) PaySim
- What it is: Synthetic mobile money transaction data
- Strengths: Has transaction types, account balances, and labels
- Limitations: Synthetic, so patterns may not reflect real fraud behavior
- Best for: Prototyping, end-to-end pipeline testing
4) BankSim
- What it is: Synthetic bank payment simulation
- Strengths: Includes customer and merchant behavior over time
- Limitations: Synthetic
- Best for: Sequence-based methods, behavior modeling
Datasets with more history / temporal structure
5) Elliptic Bitcoin Transaction Dataset
- What it is: Bitcoin transaction graph with illicit/licit labels
- Strengths: Great if you want graph-based fraud/illicit activity detection
- Limitations: Cryptocurrency domain, node labels not at the individual transaction level in the same way as card fraud
- Best for: Graph neural networks, anomaly detection on transaction networks
6) AMLSim
- What it is: Anti-money laundering synthetic transaction network simulator
- Strengths: Good for multi-account, multi-step suspicious activity chains
- Limitations: Synthetic, requires setup
- Best for: AML-style fraud and network pattern detection
More specialized or industry-like datasets
7) Synthetic Financial Datasets for Fraud Detection
- There are multiple public synthetic datasets on Kaggle and elsewhere modeled after card, banking, or payment systems.
- Strengths: Usually include time, amount, merchant/user IDs, and labels
- Limitations: Quality varies widely
- Best for: Engineering practice, simulation
8) Transaction datasets from competitions/research papers
- Many papers release datasets with:
- customer histories
- merchant categories
- device fingerprints
- sequences of events
- Strengths: Often closer to real fraud problems
- Limitations: Access can be limited; may require licensing or application
- Best for: Research-grade modeling
If you need “enough history,” prioritize these properties
Look for datasets that include:
- Ordered timestamps
- Multiple transactions per user/card/account
- Merchant and device identifiers
- Account balance or velocity features
- Labels attached to events
- Train/test splits that respect time rather than random splits
Best choices by use case
- Quick baseline: Kaggle Credit Card Fraud Detection
- More realistic tabular modeling: IEEE-CIS Fraud Detection
- Behavior over time: BankSim or PaySim
- Graph/network fraud: Elliptic or AMLSim
- Pipeline testing with custom fraud patterns: Synthetic datasets
Important caution
Many popular fraud datasets are:
- heavily anonymized
- synthetic
- not truly longitudinal at the user level
So if your app needs:
- user behavior history,
- recurring patterns,
- chargeback delays,
- merchant/customer relationships,
then synthetic datasets or specialized research datasets may be more useful than the classic benchmarks.
Practical recommendation
If you want a strong starting stack, I’d suggest:
- IEEE-CIS Fraud Detection for a richer tabular benchmark
- Kaggle Credit Card Fraud Detection for a simple baseline
- PaySim or BankSim for temporal behavior modeling
- Elliptic if you want graph-based fraud detection
If you want, I can also give you:
- a comparison table of these datasets,
- links and licenses,
- or a recommended dataset based on whether your app is card fraud, AML, or account takeover.