The quants can’t digest the documents released as fast as an AI can. They’d be crazy not to be working on an AI solution right now — and the biggest name in financial data already is. In March, Bloomberg published a paper on BloombergGPT, a 50-billion-parameter model trained on a 363-billion-token dataset drawn from its own financial sources, which the authors say “outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks.”
I’m not saying any of this should be a public product. Well — not until GPT-5 or 6 is just doing it automatically, with the right prompt and a PDF upload.
We can do this today by paying for a sentiment analysis feed and using our ingestion engine on the document releases.
A caution, before anyone gets too sanguine. The quants I’ve known have been profoundly smart — look at your miso soup and think they can refute Brownian motion smart. I would not choose them as opponents. Being first to read a document is a different skill from being right about it, and they have spent careers on the second.
That said, there is new tech to play with — or newish. FinBERT on Hugging Face describes itself in one line: “FinBERT is a pre-trained NLP model to analyze sentiment of financial text.” It’s BERT, further trained on a large financial corpus and fine-tuned on the Financial PhraseBank, and it hands back three labels — positive, negative, neutral. The paper behind it dates from August 2019, almost four years ago, and reported that “FinBERT outperforms state-of-the-art machine learning methods” on financial sentiment. I have no idea if this is any good — but 1.2 million downloads last month. Have enough of the quants adopted it? If they haven’t, four years is a long time for a free model to sit on the shelf in an industry that prizes speed.
The academics have started poking at the general-purpose models too. A paper posted in April asked ChatGPT to say whether a given headline was good, bad or irrelevant news for a company, and its authors reported that it beat the traditional sentiment methods.
Maybe a little dash of FinBERT, a financial release quickly moved into a vector database, and the right prompt?
I dunno. Wouldn’t be hard to try.
