Bloomberg announces its own financial business specialized AI ``BloombergGPT'', which can help financial analysts work and create financial news



Economic media Bloomberg has been successful in the consumer news field over the past 10 years, but it was originally a data company, and paid services such as '

Bloomberg Terminal ' that can obtain market data etc. in real time even at the time of article creation We provide services. Such Bloomberg has announced the AI model ` ` BloombergGPT '' trained using financial news and data.

[2303.17564] BloombergGPT: A Large Language Model for Finance
https://arxiv.org/abs/2303.17564



Introducing BloombergGPT, Bloomberg's 50-billion parameter large language model, purpose-built from scratch for finance | Press | Bloomberg LP
https://www.bloomberg.com/company/press/bloomberggpt-50-billion-parameter-llm-tuned-finance/

What if ChatGPT was trained on decades of financial news and data? BloombergGPT aims to be a domain-specific AI for business news | Nieman Journalism Lab
https://www.niemanlab.org/2023/04/what-if-chatgpt-was-trained-on-decades-of-financial-news-and-data-bloomberggpt-aims-to-be-a-domain- specific-ai-for-business-news/

On March 30, 2023, Bloomberg published a paper entitled 'BloombergGPT: A Large Language Model for Finance'. This Large Language Model (LLM) is an AI model specially trained with a wide range of financial data to support various natural language processing tasks within the financial industry.

Advances in AI based on LLM are yielding exciting results in many areas, but given the complexity and richness of terminology in the financial industry, ``there is a need for specialized AI models for the financial industry,'' the paper concludes. is claimed. BloombergGPT is the first step in applying LLM-based AI for the financial industry, including sentiment analysis, named entity recognition, news classification, question-and-answering, and other natural language processing tasks Bloomberg already performs in the financial industry. It seems to help improve

BloombergGPT will also marshal the vast amount of data available on the Bloomberg Terminal, opening up opportunities to better assist its customers, while also bringing the full potential of AI to the financial sector. is said to be possible.

Bloomberg GPT has 50 billion parameters, according to Tech At Bloomberg , Bloomberg's engineering arm. It was built from scratch using a unique combination of Bloomberg data and public datasets to support financial natural language processing tasks.




According to Bloomberg, BloombergGPT has been trained on a corpus of over 700 billion tokens or word fragments. OpenAI's GPT-3, released in 2020, was trained with approximately 500 billion tokens. Of this training data, 363 billion tokens are financial data that Bloomberg has collected so far, ``the largest financial industry-specific data set ever built''. The remaining 345 billion tokens are “generic datasets” obtained from other sources. The general-purpose dataset used for training BloombergGPT includes the large-scale corpus ' The Pile '.

BloombergGPT does not build a general-purpose LLM using a general-purpose dataset, nor does it build a small-scale LLM using financial-industry-specific data, but rather a hybrid approach. . A typical AI model covers many domains and achieves a high level of performance on a variety of tasks. However, BloombergGPT is also trained on financial industry-specific data, enabling it to achieve best-in-class results on financial-related benchmarks while maintaining competitive performance on general-purpose LLM benchmarks. that's right.

The graph below summarizes the NLP benchmark score on the “Finance-Specific” task and the NLP benchmark score on the “General-Purpose” task for BloombergGPT and other AI models. thing. Both 'Financcial Tasks (natural language processing benchmark for existing financial tasks)' and 'Bloomberg Tasks (natural language processing benchmark for financial tasks performed inside Bloomberg)' have given higher scores than competing AI models, and general-purpose Even in the task benchmark, you can see that both 'MMLU (large-scale multitasking language model benchmark)' and 'Reading Comprehension' score higher than the competition.



BloombergGPT has 50 billion parameters, but it has the highest level of performance compared to AI models with parameters of the same size. Also, even when compared to AI models with more parameters, BloombergGPT seems to outperform in some tasks.

BloombergGPT is trained on the same dataset as other LLMs, so it can perform the tasks you can expect from ChatGPT and others. However, you can also perform tasks that are more closely related to Bloomberg's needs, such as creating Bloomberg news article-like titles.

Below is an example of actually creating a Bloomberg news article-style title using BloombergGPT.

Input: Google has been sued by the United States and eight states to break up its ad tech business for dominating the digital advertising market. The lawsuit marks the Biden administration's first major challenge to a tech giant and is one of the rare cases since 1982 in which the Justice Department has sought to break up a major company.

Output: Google sued for monopoly of online advertising market

According to Van Spina, who has been working on FinTech development, BloombergGPT seems to have the potential to replace analysts. BloombergGPT is basically a chat-based tool that helps financial staff collect, organize, and output data. The basic financial workflow is repetitive, so it is very compatible with chat-based BloombergGPT. There are also advantages.




“The most interesting AI-related news for me this week was Bloomberg’s 50 billion parameter model trained on financial data,” said Matt Turk, who works at venture capital firm FirstMark. Companies are looking to win in AI,' he tweeted.




BloombergGPT's goal is to become the best-in-class AI model for financial tasks.

“In the long term, smaller publishers, especially those with large digitized archives, appear to be paving the way for AI models like BloombergGPT,” said the Nieman Journalism Lab. 'Of course, it's on a radically different scale than what Bloomberg creates, and could serve more as an internal tool than a public tool, but given the astounding pace of AI advancement over the past year, It may become a valuable idea sooner than you think.'

Related Posts:

in Software, Posted by logu_ii