Appendix C. Dataset and feature dictionaries

The appendix brings together the data foundations behind the book’s AI systems, emphasizing that each financial use case depends on a different kind of source and schema. It explains that credit risk models rely on large anonymized application data, fraud detection uses high-velocity transaction records, market intelligence combines pricing data with news and policy text, and personalization models are built from synthetic behavioral logs designed to preserve privacy while still capturing realistic user patterns.

For credit risk, the dataset is the AMEX Default Prediction collection, sampled down for efficiency, with features centered on customer identifiers, the binary default target, aggregated payment and delinquency measures, and anonymized categorical account-state variables. For fraud detection, the core schema includes the class label, transaction time, amount, and PCA-transformed variables that protect sensitive details while supporting supervised anomaly and fraud modeling. Together, these dictionaries show how the book structures predictive features for classification, stability, and risk scoring tasks.

The market and synthetic banking sections extend the same logic to more dynamic and behavior-driven settings. Market features blend ticker identity, adjusted prices, volatility, news source metadata, and LLM-derived sentiment or supply-risk scores to support hybrid quantitative and qualitative investing engines. The synthetic retail banking dataset uses behavioral ratios, wealth and volatility measures, clustering labels, and churn probabilities to power personalization, segmentation, and retention workflows, showing how engineered features can translate raw activity into actionable financial intelligence.