FRM Exam Part II · Advances in Artificial Intelligence: Implications for Capital Markets Activities

Machine Learning Basics for Finance in FRM Part II

Updated 11 October 2026 · Fact-checked

Machine learning lets models learn patterns from data instead of following fixed, hand-written equations. Supervised learning uses labelled outcomes, unsupervised learning finds structure without labels, and reinforcement learning learns from rewards. Deep learning uses multi-layer neural networks. Generative AI creates new content. To solve questions, match the task and data to the method, then check risks.

Understand AI and Machine Learning Fundamentals in Finance

A traditional quantitative model starts with a theory. You choose the form, such as a normal distribution for returns or a linear regression, and then estimate a few parameters. Machine learning (ML) reverses this. You give the algorithm data and a goal, and it finds the pattern itself, often with a flexible form that has many parameters.

There are three main learning types. In supervised learning, each training example has a known answer (a label). The model learns to predict it. Examples: predicting loan default (classification) or next-day volatility (regression). In unsupervised learning, there are no labels. The model finds structure, such as clusters of similar clients or trades, or unusual transactions that may be fraud or market abuse. In reinforcement learning (RL), an agent takes actions in an environment and receives rewards. It learns a policy that maximises cumulative reward. Trade execution and dynamic hedging are typical uses.

Deep learning is not a fourth learning type. It is a family of models: neural networks with many layers. Deep networks can be trained in supervised, unsupervised or reinforcement settings. This is why 'reinforcement learning vs deep learning' is a false contrast. Deep reinforcement learning combines both. Deep networks handle complex inputs such as text, images and long time series, but they are harder to explain.

Generative AI produces new content, such as text, code or synthetic data. Large language models (LLMs) are the common example. They are trained on huge text datasets to predict the next token. In capital markets they summarise research, draft reports, help write code and generate synthetic data. They can also produce confident but wrong output, called hallucination.

The key differences from traditional models are flexibility, data hunger and opacity. ML can capture nonlinear patterns, but it can overfit, which means it fits noise in the training data and fails on new data. It also needs good data and may be hard to explain. For the exam, always link the method to its risks: model risk, explainability, data quality, bias and instability when markets change.

Key formulas to remember

Supervised learning: Inputs (features) + known labels → predict label for new data; Use for prediction tasks: default, price direction, credit score. Needs labelled history.
Unsupervised learning: Inputs only, no labels → clusters, patterns, anomalies; Use for segmentation, anomaly detection, dimension reduction. No single 'right answer' to check against.
Reinforcement learning: Agent: state → action → reward → updated policy; Goal is to maximise cumulative reward over time. Suits sequential decisions such as execution.
Overfitting test: Training error low, out-of-sample error high → overfit; Check performance on data the model has not seen. Complex models overfit more easily.
Deep learning: Neural network with many hidden layers; A model class, not a learning type. Can be used within supervised, unsupervised or RL setups.

How to solve AI and Machine Learning Fundamentals in Finance questions

Use this sequence for any question on AI and ML fundamentals in finance.

1Identify the task: predicting a known outcome, finding groups or anomalies, making sequential decisions, or generating content.
2Check the data: are labels available? Is there feedback in the form of rewards? Is the input text, images or numbers?
3Map to the method: labels mean supervised; no labels mean unsupervised; rewards over sequential actions mean reinforcement; new content means generative AI.
4Decide whether deep learning is relevant: it applies when data is complex or large and the model has many layers. Treat it as a tool, not a separate learning type.
5Compare with a traditional model: note flexibility, data needs and explainability.
6Name the main risk: overfitting, poor data quality, bias, lack of explainability, hallucination or model drift.
7Choose the option that fits both the task and the risk, and reject options that confuse the categories.

Quickest way: Labels, rewards, or new content?

When to use it: Use for definition and use-case matching questions when time is short.

Ask: is there a known answer for each training example? If yes, supervised.
If no answer but you want groups or outliers, unsupervised.
If an agent acts and gets rewards, reinforcement.
If the output is new text, code or data, generative AI.
If the option says 'deep learning' as a separate category from these, be suspicious.
Pick the answer that also flags explainability or overfitting when risk is asked.

Common mistakes in AI and Machine Learning Fundamentals in Finance

Treating deep learning as a fourth learning type alongside supervised, unsupervised and reinforcement.
The terms are often listed together in news and vendor material.
Fix: Remember that learning type describes how the model is trained; deep learning describes the model structure. A deep network can be used in any of the three.
Calling anomaly detection or clustering supervised learning.
Fraud detection sounds like prediction, so students assume labels exist.
Fix: Check whether labelled fraud cases are used. If the model only looks for unusual patterns without labels, it is unsupervised.
Assuming a model with a very high in-sample fit is better.
Students equate fit with quality, as in a regression R².
Fix: Look at out-of-sample performance. A large gap between training and test results signals overfitting.
Believing generative AI outputs are factual because they sound fluent.
LLMs write confidently and in a professional tone.
Fix: Recall that LLMs predict likely text, not verified truth. Hallucination requires human review and controls.
Saying ML models replace the need for model governance.
Automation is seen as reducing human judgement.
Fix: ML adds model risk because of opacity and data dependence. Validation, monitoring and explainability still apply.
Treating reinforcement learning as learning from labelled examples.
Both involve feedback, so they seem similar.
Fix: In RL the feedback is a reward after actions, not a correct answer for each input.

Worked examples

Example 1

A bank has ten years of loan records, each marked 'defaulted' or 'did not default'. It wants a model to estimate default likelihood for new applicants. Which approach fits best, and what is the main validation concern?
A. Unsupervised clustering; concern is the number of clusters
B. Supervised classification; concern is overfitting to the training data
C. Reinforcement learning; concern is the reward function
D. Generative AI; concern is hallucination

Show the solution

Task: predict a known outcome, default or not, for new cases.
Data: historical records have labels. This points to supervised learning.
The task is classification because the output is a category or probability of a category.
Option A uses no labels, so it wastes the available information. Option C needs sequential actions and rewards, which are absent. Option D generates content, not a default estimate.
Main risk for a flexible supervised model is overfitting, so out-of-sample testing is the key check.

Answer: B

Example 2

A risk head says: 'We will use deep learning rather than reinforcement learning to build our trade execution tool.' What is wrong with this statement, and how could the two be combined?

Show the solution

Deep learning is a model structure: a neural network with many layers.
Reinforcement learning is a training approach in which an agent learns actions from rewards.
They are not alternatives on the same level, so the contrast is flawed.
Trade execution is a sequential decision problem: split an order over time, observe market impact, and adjust. This suits reinforcement learning.
A deep neural network can represent the agent's policy when market states are complex. This is called deep reinforcement learning.
Governance point: the combined model is hard to explain, so it needs testing across market conditions, limits and monitoring.

Answer: The statement confuses a model class with a learning type. Execution is naturally a reinforcement learning problem, and deep neural networks can be used inside it as deep reinforcement learning, with strong validation because of limited explainability.

Exam tips

Questions are usually scenario-based. Identify the data and goal first, then name the method.
Expect distractors that swap unsupervised and supervised, or treat deep learning as a separate learning type.
When risk is asked, the usual best answers are overfitting, explainability, data quality, bias and model drift.
For generative AI, link use cases such as summarising and coding to the hallucination and data confidentiality risks.
Compare with traditional models in terms of flexibility versus transparency, not 'better versus worse'.

Practice questions from Advances in Artificial Intelligence: Implications for Capital Markets Activities

AI and Machine Learning Fundamentals in Finance in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

AI and Machine Learning Fundamentals in Finance: frequently asked questions

What is the difference between supervised and unsupervised learning in finance?

Supervised learning trains on data with known outcomes, such as past defaults, to predict outcomes for new cases. Unsupervised learning has no labels and looks for structure, such as client clusters or unusual transactions. Pick based on whether you have labelled outcomes.

Is deep learning the same as reinforcement learning?

No. Deep learning means neural networks with many layers. Reinforcement learning means learning actions from rewards. They can be combined as deep reinforcement learning, but one describes the model and the other describes the training method.

What is generative AI in capital markets?

It is AI that creates new content, such as text, code or synthetic data. Typical uses are summarising research, drafting reports and assisting developers. Its main risks are hallucination, data leakage and weak explainability.

How do machine learning models differ from traditional quantitative models?

Traditional models usually rest on a stated theory and few parameters. ML models learn flexible patterns from data and can fit nonlinear relationships. The trade-offs are higher overfitting risk, heavy data needs and lower transparency.