17.9 Finance
Financial services have used machine learning at scale longer than most industries. Algorithmic trading firms, Renaissance Technologies, Citadel, Two Sigma, D. E. Shaw, Jane Street, have employed quantitative methods including machine learning for decades. Statistical arbitrage, market making, execution algorithms (VWAP, TWAP, implementation shortfall) and portfolio optimisation all use learned models.
The latest wave includes:
- Credit scoring: Gradient-boosted decision trees (XGBoost, LightGBM, CatBoost) remain the workhorse for credit-risk modelling. Deep learning has displaced them only in narrow areas (e.g. processing free-text or image inputs). Equifax, FICO, Experian and the various consumer-credit fintechs (Affirm, Klarna, SoFi, Upstart) all build their core risk models around gradient boosting. The Equal Credit Opportunity Act (ECOA), Fair Credit Reporting Act (FCRA) and equivalent EU and UK regulations require that adverse decisions be explainable; this has pushed the field toward interpretability tooling (SHAP values for tree ensembles is the de facto standard) rather than less interpretable deep models.
- Fraud detection: Real-time fraud detection on card transactions runs at millisecond latency on every transaction every major card network processes, tens of thousands of transactions per second. Visa's VAA (Visa Advanced Authorization) system and Mastercard's Decision Intelligence both use deep neural networks on transaction graphs and are reported to flag billions of dollars of fraud annually.
- Anti-money laundering (AML): Banks must screen transactions against sanctions lists, monitor for suspicious patterns and file Suspicious Activity Reports (SARs). Traditional rules-based AML systems produce false-positive rates above 95%. ML-based AML, including HSBC's Dynamic Risk Assessment, JPMorgan's COiN platform and various offerings from Quantexa, Featurespace and Symphony, substantially reduce false positives but raise their own questions about model auditability and regulatory acceptability.
- Quantitative trading: Hedge funds use ML for signal generation, regime detection and portfolio construction. The Renaissance Technologies Medallion Fund, returning 39% net annualised since 1988, is the canonical example, though Renaissance famously discloses essentially nothing about its methods. More recent entrants including Voleon, Two Sigma and AHL have published more openly on machine-learning-based strategies.
- Robo-advice: Betterment, Wealthfront, Nutmeg (UK) and the various platforms in this space use relatively classical portfolio-optimisation algorithms (mean-variance with Black-Litterman views) rather than the heavy ML deployed in trading and fraud. The customer-facing innovation is the user experience and tax-loss-harvesting automation rather than the optimisation method.
Generative AI has begun to enter financial workflows. Bloomberg released BloombergGPT (2023), a 50-billion-parameter LLM trained on 363 billion tokens of financial data including filings, news and Bloomberg's own data. JPMorgan's IndexGPT was announced in 2023 as a thematic investing index generator. Internal tooling for analyst report drafting, regulatory filing assistance and customer support is widely deployed though rarely externally visible.