Bloomberg’s version of ChatGPT has arrived.
On Friday, Bloomberg announced that it had launched a large-scale artificial intelligence model that has been trained on financial data to improve preexisting natural language processing tasks, such as sentiment analysis and news classification.
BloombergGPT is the first large language model designed specifically for the finance industry. It could, however, be an indicator of what’s to come in fintech: Financial firms using their own data to create ChatGPT-like models.
According to Gideon Mann, Bloomberg’s head of machine learning product and research, investors “need the ability to deal with the data deluge” they’ve been facing in recent years. BloombergGPT aims to solve that by making the query process on the company’s terminal simpler — while producing better results.
Unlike the broadly known ChatGPT technology, BloombergGPT does not have a public-facing chatbot interface. In other words, there is no BloombergGPT website into which you can type a query such as “write an investor letter” and then receive a relatively believable result.
Instead, BloombergGPT is being integrated into Bloomberg Terminal, improving the ability of users to find the information they’re seeking from the platform.
“Most of the things that we’re doing with BloombergGPT are going to be behind the scenes,” said Mann. “You’re not going to interact with that. It’s going to enhance and augment existing terminal technology.”
To build BloombergGPT, the company pulled information from its own financial data archives and from public sources to create a dataset of more than 700 billion “tokens,” or fragments of text, according to a research paper about the process used to build the model. The company then took part of this dataset and trained a “decoder-only casual language model,” which is typically used in chat-based models.
Bloomberg then tested that model with a variety of natural language processing tasks, such as reading comprehension and knowledge assessments. In a comparison against other large language models, including GPT NeoX, OPT, and BLOOM-176B, BloombergGPT outperformed on finance-related tasks.
This included, for instance, sentiment analysis. Bloomberg tested three types of analysis: equity news sentiment, equity social media sentiment, and equity transcript sentiment. On average, BloombergGPT scored 62.47 points in these categories. OPT, the large language model that scored closest to BloombergGPT, scored an average of 35.76 points.
The company is already using BloombergGPT to create “silver data,” information that is used to train other, smaller artificial intelligence models.
BloombergGPT will also help Bloomberg Terminal users find better ways to seek the information they’re looking for. The terminal uses a search language called Bloomberg Query Language to give users the ability to analyze and search for data.
But the language can be complicated for some users. “People who are using it — even if they know it, they’re not using the full power of the language,” Mann said.
The integration of BloombergGPT could allow users to obtain search results similar to what they would find using BQL, but without having to use the coding language. Instead, they can simply type in the request using plain English.
“We are moving from a world where people have to learn how to talk to a computer to a world where a computer speaks our language and can understand language much more naturally,” Mann said.
