Prompt
What dataset should I use for training a fraud detection model?
Latest observation
Oct 13, 2026
OpenAI APIWeb search: off
It depends on what kind of fraud you want to detect, but here are good dataset choices by use case:
Common public fraud datasets
-
Credit card fraud
- Kaggle Credit Card Fraud Detection
- Very popular, highly imbalanced, good for benchmarking anomaly/fraud classification methods.
- Features are PCA-transformed, so it’s good for modeling but not for interpretability.
-
Insurance fraud
- Car Insurance Fraud Detection datasets on Kaggle / open datasets from insurers or competitions.
- Useful if your target is claim fraud.
-
Banking / transaction fraud
- IEEE-CIS Fraud Detection (Kaggle)
- Great for realistic e-commerce transaction fraud; includes rich transaction and identity features.
-
Synthetic but realistic transaction fraud
- PaySim
- Simulated mobile money transaction dataset, good for prototyping and teaching.
-
E-commerce / payment fraud
- Datasets from competitions like IEEE-CIS or other Kaggle fraud competitions.
- Often more realistic than simple credit card datasets.
Which one should you choose?
- For a first model / experimentation: Kaggle Credit Card Fraud or PaySim
- For realistic feature engineering: IEEE-CIS Fraud Detection
- For domain-specific work: pick a dataset that matches your actual fraud type as closely as possible
Important note
Fraud detection performance depends heavily on:
- class imbalance
- leakage prevention
- temporal split vs random split
- cost of false positives vs false negatives
My recommendation
If you’re unsure, start with:
- IEEE-CIS Fraud Detection for a realistic challenge
- Kaggle Credit Card Fraud Detection for a simpler baseline
If you want, I can also recommend a dataset based on your specific fraud domain: credit card, insurance, banking, e-commerce, or telecom.