Lay out how AI and machine learning (ML) are used in capital markets today, and where more advanced models such as GenAI could be applied in future.
The idea
AI and machine learning are already embedded in capital markets in narrow, well-defined tasks. Typical uses are algorithmic trading and optimal execution, market surveillance that flags abnormal order patterns, and credit and market risk analytics. Others are compliance and anti-money-laundering screening, natural language processing of news and filings for sentiment signals, robo-advice, and back-office automation. Most of these are conventional ML. Supervised models predict a labelled outcome, unsupervised models find clusters or outliers, and reinforcement learning learns an action policy such as an execution schedule. Generative AI, built on large language models, creates new text, code or data. Its likely future uses are drafting research, summarising documents, writing code and answering client queries, mostly with a human reviewing the output. The usual mistake is treating all AI as GenAI. For the exam, separate established ML uses from emerging GenAI ones, and remember that hallucination and opacity limit autonomous use.
A practice question
Which of the following is an AI or machine learning use that is already established in capital markets today?
- Market surveillance models that flag abnormal trading patterns - correct
- Autonomous setting of central bank policy rates by a reinforcement learning agentWhy not: Confuses capital markets activities with monetary policy; no such autonomous use exists.
- Replacement of statutory audit opinions by a large language model without human sign-offWhy not: Overstates current GenAI use; outputs still need human review and are not a replacement for audit.
- Fully autonomous exchange rule-making by a generative modelWhy not: Invents a use case; rule-making stays with human regulators and exchanges.
Surveillance and anomaly detection, along with execution optimisation, risk analytics and compliance screening, are established ML uses. The other options describe autonomous roles that no supervisor or firm has adopted.
A harder one
A head of trading proposes giving a large language model direct authority to send orders with no human oversight. Which is the strongest reason why GenAI is more likely to be used first for internal drafting and coding support?
- GenAI cannot process text or code, so it is suited only to numerical forecasting and pricing tasksWhy not: Reverses its strength; GenAI is built to handle language and code.
- GenAI models are always slower than human traders, which makes them unusable in every capital markets taskWhy not: No blanket prohibition exists; algorithmic and ML-based execution is already common.
- Regulations explicitly prohibit any use of machine learning in order execution by regulated firms in all marketsWhy not: Overgeneralises; the limit is reliability and explainability, not a universal lack of speed.
- Plausible but wrong outputs are hard to explain, so review helps in support tasks while autonomous trading errors are costly - correct
Hallucination and opacity make unsupervised use risky where losses accrue quickly. In drafting and coding support, a reviewer catches mistakes before they have an effect, so adoption starts there.
The hardest tier
A trading firm runs three tools: (i) a gradient-boosted model trained on past fills to predict whether an order will be adversely selected; (ii) a clustering algorithm that groups trader activity to spot unusual behaviour without any labelled cases; (iii) a language model that drafts client commentary. Which classification is correct?
- (i) unsupervised ML, (ii) supervised ML, (iii) GenAIWhy not: Swaps the two ML types; a model trained on labelled outcomes is supervised and clustering without labels is unsupervised.
- (i) supervised ML, (ii) unsupervised ML, (iii) GenAI - correct
- (i) supervised ML, (ii) supervised ML, (iii) reinforcement learningWhy not: Treats clustering as supervised and a text generator as reinforcement learning; there are no labels in (ii).
- (i), (ii) and (iii) are all GenAI because each uses a neural or tree-based modelWhy not: Treats any ML as GenAI; only (iii) generates new content.
A model fitted to labelled past outcomes is supervised. Clustering without labels is unsupervised. A model that produces text is generative AI. Not all ML is GenAI.
In the app: 6 practice questions on this outcome, spaced review that brings it back before you forget it.
Start free →