Prompt
What dataset should I use for building a fraud model with transaction history?
Latest observation
Oct 16, 2026
OpenAI APIWeb search: off
If you’re building a fraud model from transaction history, the best dataset is usually one that matches your real business setting as closely as possible.
Best choice: your own transaction data
Use your own historical transactions if you have them, because fraud patterns are highly domain-specific. You’ll want:
- Transaction records
- A label for each transaction: fraud / not fraud
- Timestamps
- Customer/account identifiers
- Merchant/vendor info
- Amount, currency, channel, device, location, etc.
If you need a public dataset to start
Good public options include:
-
Kaggle Credit Card Fraud Detection
- Very popular benchmark
- Credit card transactions with fraud labels
- Highly imbalanced
- Good for testing anomaly/fraud detection approaches
- Limitation: features are anonymized, so it’s less realistic for feature engineering
-
IEEE-CIS Fraud Detection
- More realistic e-commerce transaction and identity data
- Richer feature set
- Better if you want to practice with complex tabular fraud modeling
- Still a competition dataset, so it’s not perfect for production modeling
-
PaySim
- Simulated mobile money transaction dataset
- Good for sequential/transaction-pattern fraud research
- Useful if your problem is similar to payment transfers
- Limitation: synthetic data
How to choose
Pick a dataset based on your fraud type:
- Card fraud → Kaggle Credit Card Fraud Dataset
- E-commerce fraud → IEEE-CIS Fraud Detection
- Payment transfer/mobile money fraud → PaySim
- Insurance/claims fraud → a claims-specific dataset
- Account takeover / identity fraud → dataset with device, login, and identity signals
Important features for transaction-history fraud models
Look for:
- Time-based history per user/card/account
- Transaction amount trends
- Frequency features
- Velocity features, e.g. number of transactions in last hour/day
- Geolocation changes
- Merchant/category patterns
- Device/IP changes
- Chargeback or post-event labels
Recommendation
If you’re just getting started:
- Use Kaggle Credit Card Fraud Detection for a simple baseline
- Use IEEE-CIS if you want a more realistic tabular fraud problem
- Move to your own historical transaction data for any serious model
If you want, I can also suggest:
- a dataset schema for transaction-history fraud modeling, or
- a feature engineering plan for fraud detection.