AI in Finance
AI in finance is the use of statistical and machine-learning models to make or inform decisions in banking, lending, payments, insurance and markets. Finance adopted predictive modeling early because decisions are frequent, outcomes are measurable and data is abundant, and it is also among the most regulated fields, so questions of explanation, fairness and accountability are built into how models may be used.
Fraud detection is the longest-running success. Card networks must judge each transaction in milliseconds, and the base rate of fraud is very low, so models are trained on heavily imbalanced data and tuned to trade missed fraud against false declines that annoy honest customers. HNC Software's Falcon, introduced in the early 1990s, applied neural networks to card transactions, and similar systems now combine gradient-boosted trees, behavioral profiles of each customer, device signals and graph analysis of linked accounts. Fraudsters adapt, so models must be retrained and monitored for drift, and a confirmed-fraud label can arrive weeks after the event.
Credit scoring shows the tension between accuracy and explanation. Scorecards from Fair Isaac, whose FICO score dates to 1989, were logistic-regression-style models with a small number of understandable factors. Lenders now experiment with gradient-boosted trees and neural networks, sometimes using alternative data such as bank-account cash flow, which can find creditworthy people with thin files and can also encode proxies for protected traits. In the United States the Equal Credit Opportunity Act of 1974 and Regulation B require that a declined applicant receive a statement of specific reasons. In 2022 the Consumer Financial Protection Bureau issued guidance stating that a lender using a complex algorithm must still provide accurate, specific reasons and cannot cite a checklist that does not reflect the actual model. Post-hoc explanation methods such as SHAP, from Lundberg and Lee (2017), help, but an explanation approximates the model, and critics note that interpretable models are preferable where stakes are high.
Algorithmic trading is automation of execution and strategy. Electronic markets let firms send orders in microseconds, and machine learning is applied to forecasting, execution scheduling and market making. Speed and automation create systemic hazards. During the May 6, 2010 "flash crash," US equity indexes fell sharply and recovered within minutes, and in August 2012 a faulty software deployment at Knight Capital sent unintended orders that cost the firm roughly 440 million dollars in about forty-five minutes. Predictive edges in markets are small and decay as others exploit them, so backtest results that ignore trading costs or reuse data across many trials routinely overstate what live trading delivers.
Governance is part of the technology. In the United States, the Federal Reserve and the Office of the Comptroller of the Currency issued model risk management guidance (SR 11-7) in 2011, requiring independent validation, documentation and ongoing monitoring of models that affect decisions. The European Union's AI Act classifies systems that evaluate the creditworthiness of individuals as high-risk, with requirements for data quality, oversight and documentation. Fairness is assessed through disparate impact testing and the search for less discriminatory alternative models. A well-known public controversy came in 2019, when allegations that the Apple Card gave women lower credit limits prompted a New York regulator's inquiry, which later reported no unlawful discrimination but criticized communication and transparency.
Misconceptions. Removing a protected attribute such as race or sex from the inputs does not make a model fair, because other variables can stand in for it. A more accurate model is not automatically a better one if it cannot be explained to the customer or defended to a supervisor. Related entries cover bias, explainability and anomaly detection; in finance these meet legal duties that determine which models may be deployed at all.
Read AI in finance as prediction under supervision: models earn their place by scoring risk and spotting fraud faster than people can, but they are allowed to act only to the extent that their decisions can be explained, tested for fairness and held to account.