Bloomberg is developing an LLM (large language model) fed with hundreds of billions of financial and economic data. Named BloombergGPT, this generative AI will help the economic information group with a multitude of tasks.

Bloomberg is developing its own LLM (large language model). In its announcement, the business intelligence group says its generative AI, dubbed BloombergGPT, outperforms comparable models on language processing tasks in finance.

A model trained with hundreds of billions of data

Bloomberg explains that the particular vocabulary and complexity of the financial field requires a domain-specific model. In a research paper, researchers from the company and Johns Hopkins University explain that to develop the first version of their model, it was fed with a mass of impressive data (financial documents, news, filings, press releases, social media) from Bloomberg archives.

The group’s financial information gathering activities, initiated four decades ago, have allowed it to feed the generative AI with a set of 363 billion tokens (basic units of text or code that an AI LLM uses to process and generate language). To which are added public data sets, to ultimately create a vast training corpus comprising more than 700 billion tokens. The ultra-specialized model thus created has no less than 50 billion parameters (for comparison, GPT-4 would have 1000 billion, a figure not confirmed by OpenIA).

Categorization of news and responses to queries

The financial news firm says BloombergGPT will help with things like sentiment analysis, named entity identification, news categorization and query response. “BloombergGPT will allow us to tackle many new types of applications, while delivering much better out-of-the-box performance than custom models for each application, with a faster time to market,” says Shawn Edwards, chief technology officer at Bloomberg.

Following a series of tests, created internally but also publicly available benchmarks, Bloomberg claims that its homemade GPT model “significantly outperforms existing open models of similar size in financial tasks, while achieving equal or superior in general NLP benchmark tests”.