{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreibd3i2pqyr32j5mnbv7adekjzy2dcxbvuayycncq22sgf4ubsc74m",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mk7mrvbh6bl2"
  },
  "path": "/t/are-there-any-good-open-datasets-for-training-fintech-models-fraud-detection-credit-scoring-etc/175482#post_2",
  "publishedAt": "2026-04-24T01:56:33.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "Kaggle",
    "Hugging Face",
    "GitHub",
    "UCI Machine Learning Repository",
    "OpenML",
    "docs.interpretable.ai",
    "Consumer Financial Protection Bureau",
    "The Securities and Trade Commission.",
    "FRED",
    "World Bank Data Help Desktop"
  ],
  "textContent": "Seems so many…\n\n* * *\n\nThere are **good free/open datasets for fintech AI** , but they are best used for **learning, prototyping, research, benchmarking, and demos** — not as a complete replacement for real bank, lender, or payment-company data.\n\nThe easiest way to think about it:\n\nGoal | Good data type\n---|---\nFraud detection | Card transactions, bank transfers, synthetic fraud data\nCredit scoring | Loan/default datasets, mortgage data, credit-risk benchmarks\nAML / money laundering | Synthetic bank-transfer data, crypto transaction graphs\nFinance chatbots / RAG | SEC filings, financial reports, complaints, financial news\nMarket/macro apps | Economic time series, public financial indicators\n\n* * *\n\n# Best open datasets by use case\n\n## 1. Fraud detection\n\n### **Credit Card Fraud Detection — Kaggle / ULB**\n\nThis is the classic beginner dataset for payment-card fraud detection. It contains **284,807 transactions** and **492 fraud cases** , so it is highly imbalanced. That makes it useful for learning why fraud detection is hard: fraud is rare, and simple accuracy can be misleading. (Kaggle)\n\n**Good for:**\n\n  * beginner fraud models\n  * imbalanced classification\n  * anomaly detection\n  * precision/recall practice\n  * threshold tuning\n\n\n\n**Main caution:**\nMost features are anonymized, so it is not very business-interpretable. It is good for learning, not for proving a real fraud system is production-ready.\n\n* * *\n\n### **Bank Account Fraud Dataset Suite, BAF**\n\nBAF is a public suite of **synthetic bank-account-opening fraud datasets**. It was published with NeurIPS 2022 and was designed to capture realistic problems such as **class imbalance, bias, and time dynamics**. (Kaggle)\n\n**Good for:**\n\n  * account-opening fraud\n  * fairness testing\n  * synthetic fraud modeling\n  * temporal validation\n  * tabular ML benchmarks\n\n\n\n**Main caution:**\nIt is synthetic. That is useful for privacy, but a model can learn patterns from the generator rather than real fraud behavior.\n\n* * *\n\n### **Synthetic Financial Datasets for Fraud Detection — Hugging Face**\n\nThis Hugging Face dataset is available in **Parquet** format and has around **millions of rows** , making it convenient for larger fraud-detection experiments. (Hugging Face)\n\n**Good for:**\n\n  * scalable fraud modeling\n  * tabular ML\n  * pipeline testing\n  * Hugging Face / pandas workflows\n\n\n\n**Main caution:**\nSynthetic fraud data is useful for practice, but it should not be treated as real bank validation.\n\n* * *\n\n## 2. AML / money-laundering detection\n\n### **IBM AML-Data**\n\nIBM’s AML-Data repo provides synthetic financial transactions such as bank transfers, purchases, credit-card transactions, and checks. Most transactions are legitimate, while some represent money laundering; the data is in CSV format and generated using a multi-agent virtual-world model. (GitHub)\n\n**Good for:**\n\n  * AML transaction monitoring\n  * suspicious-transaction classification\n  * graph features\n  * money-flow analysis\n  * alert-ranking demos\n\n\n\n**Main caution:**\nSynthetic AML patterns may be easier or cleaner than real laundering behavior.\n\n* * *\n\n### **IBM AMLSim**\n\nAMLSim is a multi-agent simulator for generating synthetic banking transaction data so researchers can test AML algorithms on shared synthetic data. (GitHub)\n\n**Good for:**\n\n  * generating custom AML data\n  * testing money-laundering typologies\n  * graph-based AML experiments\n  * synthetic transaction simulation\n\n\n\n**Main caution:**\nYou may need engineering work to generate exactly the scenarios you want.\n\n* * *\n\n### **Elliptic++ Bitcoin Dataset**\n\nElliptic++ is useful for crypto-related AML. It contains about **203k Bitcoin transactions** and **822k wallet addresses** , enabling illicit-transaction and illicit-address detection with graph data. (GitHub)\n\n**Good for:**\n\n  * crypto AML\n  * graph neural networks\n  * illicit transaction detection\n  * wallet/address risk scoring\n\n\n\n**Main caution:**\nBitcoin graph behavior does not automatically generalize to normal bank transfers, card payments, or lending.\n\n* * *\n\n## 3. Credit scoring and default prediction\n\n### **German Credit Data — UCI**\n\nGerman Credit is a classic credit-risk dataset. It classifies people as good or bad credit risks and has **1,000 instances** and **20 features**. (UCI Machine Learning Repository)\n\n**Good for:**\n\n  * beginner credit scoring\n  * scorecard modeling\n  * fairness demos\n  * cost-sensitive classification\n\n\n\n**Main caution:**\nIt is small and old. Use it for learning, not for production credit decisions.\n\n* * *\n\n### **Default of Credit Card Clients**\n\nThis dataset has **30,000 instances** , **24 features** , and a binary default label. It is commonly used for credit-card default prediction. (OpenML)\n\n**Good for:**\n\n  * default prediction\n  * credit-risk classification\n  * calibration\n  * fairness checks\n  * explainability demos\n\n\n\n**Main caution:**\nIt is older and jurisdiction-specific, so do not assume it reflects your customer base.\n\n* * *\n\n### **Home Credit Default Risk — Kaggle**\n\nThis Kaggle competition dataset asks whether each applicant is capable of repaying a loan. The data includes a main application table, with one row per loan, and a target for the training set. (Kaggle)\n\n**Good for:**\n\n  * more realistic credit-risk modeling\n  * feature engineering\n  * relational/tabular data\n  * loan repayment prediction\n\n\n\n**Main caution:**\nIt is more complex than beginner datasets. Also, Kaggle datasets may have specific competition/data-use rules.\n\n* * *\n\n### **FICO HELOC Dataset**\n\nThe FICO HELOC dataset is widely used for explainable credit-risk modeling. The task is to predict whether applicants will repay a home-equity line of credit within two years. (docs.interpretable.ai)\n\n**Good for:**\n\n  * explainable AI\n  * credit underwriting examples\n  * scorecards\n  * interpretable ML\n  * adverse-action-style explanations\n\n\n\n**Main caution:**\nCheck access terms before using it commercially.\n\n* * *\n\n### **HMDA Mortgage Data**\n\nHMDA is one of the most important public datasets for U.S. mortgage analysis. The CFPB describes HMDA data as the most comprehensive public source of information on the U.S. mortgage market. (Consumer Financial Protection Bureau)\n\n**Good for:**\n\n  * mortgage lending analysis\n  * fair-lending research\n  * loan approval/denial analysis\n  * geographic and demographic studies\n\n\n\n**Main caution:**\nHMDA is not a full credit-bureau dataset. It does not contain every underwriting variable a lender would use.\n\n* * *\n\n## 4. Finance NLP, chatbots, and RAG\n\n### **SEC EDGAR APIs**\n\nThe SEC provides RESTful APIs for company submissions and XBRL financial-statement data. The APIs return JSON, need no authentication or API key, and include submissions history plus XBRL data from filings such as 10-K, 10-Q, 8-K, 20-F, 40-F, and related forms. (The Securities and Trade Commission.)\n\n**Good for:**\n\n  * finance chatbots\n  * SEC filing search\n  * 10-K / 10-Q analysis\n  * financial-statement extraction\n  * RAG systems\n  * company-risk summarization\n\n\n\n**Main caution:**\nRaw filings are long and messy. You need good chunking, retrieval, citations, and date handling.\n\n* * *\n\n### **PleIAs/SEC — Hugging Face**\n\nThis Hugging Face dataset contains SEC annual reports, Form 10-K, from **1993 to 2024** , stored in **Parquet** format. (Hugging Face)\n\n**Good for:**\n\n  * SEC filing RAG\n  * finance document search\n  * long-document summarization\n  * financial text embeddings\n  * risk-factor extraction\n\n\n\n**Main caution:**\nA chatbot trained or built on filings should cite sources and avoid giving unsupported investment advice.\n\n* * *\n\n### **Financial PhraseBank**\n\nFinancial PhraseBank contains **4,840 English financial-news sentences** labeled by sentiment. (Hugging Face)\n\n**Good for:**\n\n  * financial sentiment classification\n  * small NLP baselines\n  * fine-tuning a classifier\n  * positive/neutral/negative financial text analysis\n\n\n\n**Main caution:**\nIt is small. It is better for evaluation or a simple classifier than for training a large financial language model.\n\n* * *\n\n### **CFPB Consumer Complaint Database**\n\nThe CFPB Consumer Complaint Database lets users explore, filter, map, read, and export consumer complaints about financial products and services. (Consumer Financial Protection Bureau)\n\n**Good for:**\n\n  * complaint classification\n  * customer-support routing\n  * financial product taxonomy\n  * consumer-finance NLP\n  * topic modeling\n  * trend detection\n\n\n\n**Main caution:**\nComplaints are not a random sample of all customers. They reflect people who chose to complain.\n\n* * *\n\n## 5. Market, macroeconomic, and public financial data\n\n### **FRED API**\n\nThe FRED API gives programmatic access to economic data from FRED and ALFRED. It can retrieve data by source, release, category, series, and other parameters. (FRED)\n\n**Good for:**\n\n  * interest rates\n  * inflation\n  * unemployment\n  * GDP\n  * credit-cycle indicators\n  * macroeconomic features\n\n\n\n**Main caution:**\nFRED is great for context, but it is not customer-level fintech data.\n\n* * *\n\n### **World Bank Indicators API**\n\nThe World Bank Indicators API provides programmatic access to nearly **16,000 time-series indicators** across many databases, with many series going back more than 50 years. (World Bank Data Help Desktop)\n\n**Good for:**\n\n  * country risk\n  * financial inclusion analysis\n  * macroeconomic modeling\n  * emerging-market fintech analysis\n  * development-finance applications\n\n\n\n**Main caution:**\nMost indicators are country-level or macro-level, not transaction-level.\n\n* * *\n\n# Best starting choices\n\n## If you want to build a fraud model\n\nStart with:\n\n  1. **Credit Card Fraud Detection**\n  2. **BAF**\n  3. **IBM AML-Data**\n  4. **Elliptic++** if you want crypto/graph fraud\n\n\n\nUse metrics like:\n\n  * precision\n  * recall\n  * PR-AUC\n  * precision at top-k\n  * recall at fixed false-positive rate\n\n\n\nDo **not** rely only on accuracy. Fraud is usually rare, so accuracy can look good even when the model misses most fraud.\n\n* * *\n\n## If you want to build a credit scoring model\n\nStart with:\n\n  1. **German Credit**\n  2. **Default of Credit Card Clients**\n  3. **Home Credit Default Risk**\n  4. **FICO HELOC**\n  5. **HMDA** for mortgage and fair-lending analysis\n\n\n\nFor credit scoring, also think about explainability. The CFPB has said lenders using AI or complex credit models must provide **specific and accurate reasons** when taking adverse action against consumers. (Consumer Financial Protection Bureau)\n\n* * *\n\n## If you want to build a finance chatbot or RAG app\n\nStart with:\n\n  1. **SEC EDGAR APIs**\n  2. **PleIAs/SEC on Hugging Face**\n  3. **CFPB complaints**\n  4. **Financial PhraseBank**\n  5. **FRED / World Bank indicators**\n\n\n\nThis is often a safer and more practical starting point than credit scoring because public filings and complaints are real public text data.\n\n* * *\n\n# Simple ranking: best datasets by beginner-friendliness\n\nDataset/source | Beginner-friendly? | Best use\n---|---|---\nCredit Card Fraud Detection | High | fraud basics\nGerman Credit | High | credit scoring basics\nDefault of Credit Card Clients | High | default prediction\nFinancial PhraseBank | High | sentiment classification\nCFPB complaints | Medium | finance NLP\nSEC EDGAR / PleIAs SEC | Medium | finance RAG\nHome Credit Default Risk | Medium/hard | advanced credit modeling\nIBM AML-Data | Medium | AML transaction modeling\nBAF | Medium | realistic fraud/fairness research\nElliptic++ | Hard | crypto graph ML\n\n* * *\n\n# Important warning\n\nOpen fintech datasets are useful, but they usually have limits:\n\n  * Many fraud datasets are **synthetic** or **anonymized**.\n  * Many credit datasets are **old** or **small**.\n  * Public datasets rarely contain the full data a bank or lender would use.\n  * Real fraud and AML labels are hard to publish because of privacy and security.\n  * Production credit models need legal, compliance, fairness, and explainability review.\n\n\n\nSo the best use of open fintech data is:\n\n> Build prototypes, learn modeling techniques, test pipelines, create demos, and benchmark methods.\n\nThe wrong use is:\n\n> Train on a public dataset and assume it is ready for real lending, fraud blocking, or AML decisions.\n\n* * *\n\n# Short summary\n\n  * **Yes, good free fintech datasets exist.**\n  * For **fraud** , start with Credit Card Fraud Detection, BAF, IBM AML-Data, and Elliptic++.\n  * For **credit scoring** , start with German Credit, Default of Credit Card Clients, Home Credit, FICO HELOC, and HMDA.\n  * For **finance NLP/RAG** , use SEC EDGAR, PleIAs/SEC, CFPB complaints, and Financial PhraseBank.\n  * For **macro/market context** , use FRED and World Bank indicators.\n  * These datasets are best for **learning, prototyping, research, and demos** , not direct production deployment.\n\n",
  "title": "Are there any good open datasets for training fintech models (fraud detection, credit scoring, etc.)?"
}