Large language models and generative AI in finance
A large language model (LLM) is a transformer-based neural network that learns from vast text collections and writes text: reports, summaries, sentiment scores and insights. Generative artificial intelligence (AI) creates new content or synthetic data from learned patterns. Both need human checks, because numbers are not an LLM's core skill.
Before this: AI and machine learning: supervised, unsupervised, reinforcement, deep
On the exam
- Describe how big data, machine learning, and artificial intelligence are used in financial data science, fintech, and investment management
Describe how large language models and generative AI are used in finance. Expect: what LLMs are used for (analyzing reports and generating insights, with human oversight); why their figures can be wrong; how they are trained; the advantage of generative AI (synthetic data from learned patterns, still built on historical data and with no guarantee of accuracy); and why it is not Monte Carlo simulation or bootstrapping. No calculations.
How deeply to learn each part
- Know the idea
- Understand it in general. It frames other topics, but you will not calculate with it.
- Use it
- Understand it well enough to solve a question with it. Recognize when it applies.
- Memorize
- Learn it by heart. The exam will not give it to you.
You know this one. The full lesson is folded away. The takeaways, formulas and checkpoint below are your quick refresher, and the checkpoint is a fair test of whether you still do.
You tested strong here. The full lesson is folded away. The takeaways, formulas and checkpoint below are your quick refresher, and the checkpoint is a fair test of whether you still know it.
What is a large language model?
A large language model (LLM) , such as a generative pre-trained transformer (GPT) , is a neural network trained on vast amounts of text that can read and write language. It is one of the biggest recent advances in artificial intelligence (AI).
It is built on a transformer , a network design that handles long-range dependencies : how words far apart in a text relate to each other, which is what gives text its context and meaning. The model picks out the most relevant parts of an instruction or document and, drawing on what it learned, produces whole sentences and longer passages that fit the context, not just a likely next word.
In finance, LLMs analyze company reports, economic indicators, past performance data and financial news, combining financial knowledge with general knowledge.
How is a financial LLM trained?
In stages, mostly without hand-labeled examples.
- General language self-supervised; grammar, meaning
- Specialized financial corpora Use it reports, filings, call transcripts
- Fine-tuning or transfer learning Use it adapt to a task
- Uses Memorize
- Text generation
- Summarization
- Sentiment analysis
- Insight generation
-
- General language. The model first reads a huge body of general text to learn grammar, meaning and the patterns of language. It uses self-supervised learning : it learns from the structure of the data itself (for example, the text around a word), not from labels that people attach. This is not supervised learning: nobody labels the examples.
- Specialized corpora. Next it reads specialized corpora : collections of financial texts such as annual reports, disclosure statements, earnings-call transcripts and market commentary. These teach the vocabulary and nuance of finance.
- Adaptation. Fine-tuning or transfer learning then adapts the trained model to a specific task, such as summarizing credit reports.
What do LLMs do in finance?
There are four main LLM uses , and all four are worth memorizing:
- Text generation: human-like reports and investment summaries.
- Summarization: long reports or data condensed for quicker decisions.
- Sentiment analysis: the tone of news or social media, to anticipate market moves.
- Insight generation: making sense of jargon and complex data for analysts and portfolio managers.
Sentiment analysis is a natural language processing (NLP) task (third lesson of this module). An LLM is a more powerful engine for it.
Where do LLMs go wrong?
An LLM is trained to predict sequences of words, so it writes fluent text, but getting financial figures right is not its core skill. Expect errors in reading or reporting metrics. This is the LLM accuracy risk .
Reports written entirely by an LLM, or decisions taken on its output without human oversight , can lead to errors in investment strategy, regulatory compliance and analysis, mislead readers and erode trust.
The fix is a hybrid approach . Combine LLM output with traditional data-driven models (for example, pull figures from the structured data, not from generated text) or have people review it as they would review an analyst’s work. This keeps the speed and catches the errors.
Exam tip If an answer option says an LLM can work without human oversight, or has replaced analysts, it is wrong.
What is generative AI?
Generative AI creates new content, data or solutions from patterns it has learned in existing data. Traditional AI models classify or predict a value (regression); generative models produce outputs that look like real data. It uses neural networks such as generative adversarial networks (GANs) and variational autoencoders (VAEs), covered in the machine learning lesson, and transformers.
One use: train a generative model on years of market data, then have it create many synthetic data scenarios, including stressed ones that never happened, and test a trading algorithm against all of them.
The advantage is that it can create synthetic data based on what it learned, saving effort no team could spare. It still needs historical data to learn from, and its scenarios are plausible, not predictions: it does not guarantee accurate forecasts.
Generative AI, Monte Carlo or bootstrapping?
All three produce scenarios, but differently. Generative AI is neither Monte Carlo simulation nor bootstrapping.
- Bootstrapping Memorize redraw past data with replacement
- Monte Carlo simulation Memorize draw from an assumed distribution
- Generative AI Memorize learn the structure, create new samples
| Method | Where scenarios come from | What it cannot do |
|---|---|---|
| Bootstrapping | redraws observed past data at random, with replacement | create a value that never occurred |
| Monte Carlo simulation | random draws from a distribution the analyst specifies (for example, normal with a chosen mean and standard deviation) | find a pattern the analyst did not assume |
| Generative AI | a model that has learned the data’s underlying structure generates new samples resembling it | guarantee the scenarios are accurate forecasts |
Bootstrapping and Monte Carlo are covered in Simulating Prices and Returns.
Memorize
Who oversees AI in finance?
Regulation of AI in finance is evolving quickly, driven by data privacy, the need for algorithms to be transparent and the risk that automated decisions spread problems through the system.
Regulated institutions (banks and related providers) must meet strict standards on data management, model validation and risk assessment. Non-regulated firms have more freedom but still owe clients contractual and legal duties of fair dealing.
Model risk is the danger from models that are poorly built, insufficiently tested or not verified. It concerns both kinds of firm, and the answer is rigorous development, testing and verification.
Using scraped or vendor data adds data-protection questions and the question of material nonpublic information (see the first lesson of this module and Standard II(A)).
Go deeper What else can AI read?
Machine learning applied to big data now helps predict events such as whether a merger succeeds or how an election turns out. Image recognition reads satellite pictures of parking lots, ports, factories and crop fields, and such signals can feed company valuations or economic models. The drivers are more data, better algorithms, faster computing and cheaper storage.
Memorize this
MemorizeLLMs in finance
| Remember | |
|---|---|
| What | transformer-based neural network trained on vast text; e.g. generative pre-trained transformer (GPT) |
| Trained by | self-supervised learning on general text, then specialized financial corpora, then fine-tuning or transfer learning |
| Used for | text generation, summarization, sentiment analysis, insight generation |
| Weak spot | accuracy of financial figures (not language, so not its core skill) |
| Fix | hybrid approach: data-driven models or human review |
Memory hook Generate, Summarize, Sense, See (four uses). Words are its skill, numbers its blind spot.
Know it cold An LLM never runs without human oversight in an exam answer.
MemorizeGenerative AI
| Method | Scenarios come from |
|---|---|
| Bootstrapping | past data redrawn with replacement |
| Monte Carlo | an assumed distribution |
| Generative AI | a learned model of the data: new samples (generative adversarial networks, variational autoencoders, transformers) |
Memory hook Bootstrap redraws, Monte Carlo assumes, generative AI learns.
Know it cold Advantage: synthetic data from learned patterns. It still needs historical data and does not guarantee accurate forecasts.
- General language self-supervised; grammar, meaning
- Specialized financial corpora Use it reports, filings, call transcripts
- Fine-tuning or transfer learning Use it adapt to a task
- Uses Memorize
- Text generation
- Summarization
- Sentiment analysis
- Insight generation
-
- Bootstrapping Memorize redraw past data with replacement
- Monte Carlo simulation Memorize draw from an assumed distribution
- Generative AI Memorize learn the structure, create new samples
How it is asked
-
How LLMs are used
Which describes the use of LLMs in finance?
They analyze reports and generate insights, with human oversight. Traps: "translate basic text only", "replace analysts".
-
Why the numbers slip
Why can an LLM-written earnings summary misstate a margin?
Its core skill is predicting words; financial figures are not language. Trap: "it was trained on labeled data".
-
Generative AI advantage
Main advantage of generative AI in scenario analysis?
Creates synthetic data from patterns learned in existing data. Traps: "needs no historical data", "guarantees accurate forecasts".
-
Generative AI vs simulation
A model learns 20 years of market behavior and creates new scenarios. Monte Carlo?
No: generative AI, it learns the structure. Monte Carlo draws from an assumed distribution; bootstrapping redraws past data.
Exam traps
- Saying LLMs can produce analysis or reports on their own, without human oversight, or have replaced analysts.
- Saying LLMs are mainly for basic translation. They generate, summarize, score sentiment and produce insights.
- Calling LLM training supervised. It is self-supervised: the model learns from the structure of the text, without manual labels.
- Saying generative AI removes the need for historical data or guarantees accurate predictions.
- Calling generative AI scenario creation Monte Carlo simulation or bootstrapping.
- Thinking only regulated banks face model risk. Every firm using models does.
Checkpoint
3 exam-style questions. About 90 seconds each.
1A research team plans to use a large language model (LLM) to draft summaries of quarterly filings for its portfolio managers. Which approach is most appropriate?
2An LLM-written earnings summary states a company's margin incorrectly even though every sentence reads well. The most likely reason is that the LLM:
3A risk team trains a model on 20 years of market data and uses it to create thousands of new crisis scenarios, unlike any single past period, to stress-test a trading algorithm. Compared with Monte Carlo simulation and bootstrapping, this generative model most likely:
Key takeaways
- An LLM is a transformer-based neural network trained on vast text, first general, then financial, with self-supervised learning, then fine-tuned.
- LLMs generate text, summarize, score sentiment and produce insights; figures are their weak spot.
- Never rely on LLM output without oversight: combine it with data-driven models or human review.
- Generative AI creates synthetic data from learned patterns; it still needs historical data, guarantees nothing, and is not Monte Carlo or bootstrapping.
- Regulators focus on data privacy, transparency and systemic risk; model risk concerns every firm.
Questions about this lesson?
Where this comes back
- Portfolio Management: An Overview Portfolio Construction
Questions candidates ask
How are large language models used in finance?
To analyze company reports, economic data and news, and to generate text, summaries, sentiment scores and insights for analysts and portfolio managers. Because accuracy with financial figures is not their core skill, their output is reviewed by people or checked against data-driven models before anyone relies on it.
How is a financial LLM trained?
By self-supervised learning, without hand-labeled examples: first on a vast body of general text to learn the language, then on specialized financial corpora such as annual reports and call transcripts. Fine-tuning or transfer learning then adapts it to a specific task.
What are the risks of using LLMs for financial reports?
Fluent text can carry wrong numbers, because LLMs are trained to predict words, not to verify data. Reports written entirely by an LLM without oversight can cause errors in strategy, compliance and analysis. A hybrid approach, with human review or data-driven checks, reduces the risk.
How is generative AI different from Monte Carlo simulation?
Monte Carlo draws scenarios from a distribution the analyst assumes, and bootstrapping redraws past data. Generative AI learns the underlying structure of the data and creates new samples that resemble it. It still needs historical data and does not guarantee accurate forecasts.
Spotted an error? Tell us and we will fix it.