ChatGPT

ChatGPT is a conversational artificial intelligence service developed by OpenAI, a company based in San Francisco. A user types a question or instruction, called a prompt, and the system replies in natural language. It can explain concepts, summarize documents, draft text, write and correct computer code, and work through data supplied to it. It was released to the public on 30 November 2022.

The service is built on a series of models called generative pre-trained transformers, from which the GPT in its name is taken. Its release brought the large language model to a mass audience and was followed by rival products from other technology companies. In finance it is used as a research and programming aid. It is a general-purpose tool and not a trading system, a data feed or a licensed adviser.

OpenAI was founded in 2015. Its GPT models are neural networks of the transformer type, trained on very large amounts of text to predict the next word in a sequence. ChatGPT placed a chat interface on top of such a model and was first offered free as a research preview. Within about two months it was estimated to have reached 100 million users, at the time among the fastest take-ups of any consumer application. A paid subscription followed in early 2023, and the GPT-4 model was released in March 2023. Microsoft is a major investor in OpenAI and has built the models into its own products.

Later versions added the ability to read images and files, browse the web, run code and call outside tools, as well as access for developers through an application programming interface. Features and model versions change often, and the current list is given on the company's own site.

The model does not look answers up in a database. It generates a reply one small unit of text at a time, each chosen according to probabilities learned in training. This is a form of deep learning applied to natural language processing. Because the output is sampled, the same prompt can give different answers on different occasions.

The quality of a reply depends heavily on how the request is worded and what context is supplied, a skill known as prompt engineering. Unless it is connected to a search tool or a data source, the model knows only what was in its training data up to a cut-off date and has no view of current prices or news.

• Research support. Analysts use it to summarize filings, earnings call transcripts, research papers and regulatory documents, and to compare them.
• Programming. Traders use it to write and debug strategy code, for example in Python or in a platform language such as Pine Script, and to explain existing code.
• Text scoring. Language models can label headlines or social media posts as positive or negative, a task known as sentiment analysis. Academic studies from 2023 onward have tested whether such scores predict short-term returns, with some reporting an effect before trading costs.
• Education. It is widely used to explain financial terms, formulas and market conventions.
• Automation. Connected to tools, the models can carry out multi-step tasks, an approach described as agentic AI.

These uses support the work around algorithmic trading. The chat service itself does not place orders, and any link to a brokerage account is built by the user or a third party.

The best-known weakness is that the model can produce statements that are fluent, confident and false, including invented figures, quotations and sources. This is often called hallucination. Numerical work is a particular risk: calculations may be wrong unless the system runs code to perform them, and financial data quoted from memory may be out of date or simply incorrect. Every fact and number it supplies needs to be checked against a primary source.

Code written by the model can contain subtle errors. In a strategy test, such errors include look-ahead bias, where the code uses information that was not available at the time of the trade. The ease of generating many strategy ideas also raises the risk of overfitting during backtesting. Studies that test the model's predictions on past news face a related problem, since the model may have seen later outcomes in its training data.

Text entered into the service is sent to OpenAI's systems. Many banks and asset managers restrict or control its use so that client data, trading positions and unpublished research are not disclosed. Firms that use language models in regulated activities remain responsible for the output, as with any other tool.

Regulators have also acted against exaggerated claims. In March 2024 the Securities and Exchange Commission announced settled charges against two investment advisers for making false or misleading statements about their use of artificial intelligence, a practice it called AI washing. ChatGPT is not authorized to give personal investment advice, and its terms place responsibility for how output is used on the user.

• OpenAI
• ChatGPT