TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

BloombergGPT: A Large Language Model for Finance

175 点作者 SerCe大约 2 年前

22 条评论

hn_throwaway_99大约 2 年前
I&#x27;m curious about how large language models will do in finance, considering the one thing LLMs do remarkably poorly is <i>math</i>.<p>I use ChatGPT to keep track of tasks and Todo lists. It works phenomenally well for me, and the natural language back-and-forth helps keep me motivated. I give it a set of tasks, with time estimates, and it organizes these tasks for me, and I tell it when I complete them, and it updates my task list.<p>The one funny mistake it makes is that when it groups my tasks (say I have 3 &quot;Work&quot; tasks and 2 &quot;Personal&quot; tasks&quot;) it sums up the total estimated time for each task group, but the totals are often wrong, especially when I start adding new tasks or completing tasks.<p>When so much of finance requires numeric accuracy, I&#x27;m curious how BloombergGPT handles numbers.
评论 #35392745 未加载
评论 #35382445 未加载
评论 #35384110 未加载
评论 #35383186 未加载
choeger大约 2 年前
I wonder if we should go back to ontologies now. It should be fairly easy to have a LLM generate entries for any ontology. Then we need to check them for truth, obviously, but that can be parralized and potentially automated.<p>Say, I&#x27;d like an ontology of the current stock market focusing on the relationship between natural persons and public companies, board members, well-known analysts, investors and so on. This would be tedious for anyone to do, but should be fairly simple with a LLM.<p>Another task, maybe a little bit further into the future is categorizing open source intelligence. Think of oryxspioenkop.com and their famous lists of lost equipment in the Russian invasion of Ukraine. It&#x27;s tedious and time-consuming but generates a valuable dataset. Here, image recognition would be necessary, but the principle is still the same, no?<p>Come to think of it, how does a company like TomTom generate map data nowadays?
wanderingmind大约 2 年前
No API, no code, no model to test, we are to believe the verbiage that the paper contains at its face value?
评论 #35392792 未加载
评论 #35382193 未加载
Michelangelo11大约 2 年前
Wait... judging by their own press release, this thing looks terrible. They didn&#x27;t benchmark it on finance-specific tasks against GPT-4, or even GPT-3.5, just a bunch of more-or-less random also-ran models. It beat those models, but that&#x27;s not saying much at all, and the fact that they didn&#x27;t benchmark it against any OpenAI model in finance-specific tasks speaks volumes.<p>(They did benchmark it against GPT-3 in general-purpose tasks and, unsurprisingly, GPT-3 came out on top.)<p>Press release link: <a href="https:&#x2F;&#x2F;www.bloomberg.com&#x2F;company&#x2F;press&#x2F;bloomberggpt-50-billion-parameter-llm-tuned-finance&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.bloomberg.com&#x2F;company&#x2F;press&#x2F;bloomberggpt-50-bill...</a>
评论 #35392680 未加载
nl大约 2 年前
Reading the paper it&#x27;s apparent what a huge contribution HuggingFace made by doing the BLOOM work in the open. There&#x27;s so much knowledge there that anyone else trying to train a LLM can use.<p>Also it looks like public filings are a large dataset that isn&#x27;t currently being used by other LLMs.
yanglet大约 2 年前
A few points to share:<p>1). Finance is high dynamic. BloombergGPT retrains LLM using a mixed dataset of finance and general sources is too much expensive (1.3M hours). Lightweight adaptation is highly favorable.<p>2). Internet-scale finance data (timely updates using an automatic data curation pipeline) is critical. BloombergGPT has privileged data access and API access. A promising alternative is &quot;democratizing Internet-scale finance data&quot;.<p>3). Another key technology is &quot;RLHF (Reinforcement learning from human feedback)&quot;, which is missing in BloombergGPT. RLHF enables learning individual preferences (risk-aversion level, investing habits, personalized robo-advisor, etc.)
geenew大约 2 年前
This must be the &#x27;big data&#x27; I heard so much about.<p>LLMs in general, I mean. They seem to be the first widespread application for large, unstructured datasets. Still hype-y, but maybe even a &#x2F;practical&#x2F; application.
评论 #35383858 未加载
vgeek大约 2 年前
What about LTCM-GPT?
yanglet大约 2 年前
Anyone interested in building a ChatGPT for FinTech, or FinGPT? I and my team are actively developing this project: <a href="https:&#x2F;&#x2F;github.com&#x2F;AI4Finance-Foundation&#x2F;ChatGPT-for-FinTech">https:&#x2F;&#x2F;github.com&#x2F;AI4Finance-Foundation&#x2F;ChatGPT-for-FinTech</a>
yanglet大约 2 年前
FinGPT better be open-source and open-finance.<p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=35403167" rel="nofollow">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=35403167</a>
PaulHoule大约 2 年前
i like how they used a little more than 50% of domain specific text and a little less of general text. it beats the other LLMs on financial tasks for 4 out of 5 tasks usually by a large margin and but another one squeaks past it on NER.
评论 #35381753 未加载
alecco大约 2 年前
I would&#x27;ve made InvestopediaGPT. Much better quality and better for training.
mightytravels大约 2 年前
Where is my &#x27;ArbitrageGPT&#x27;? Renaissance might be working on that...
评论 #35390385 未加载
评论 #35381593 未加载
blitzo大约 2 年前
One industry i could see in trouble from this is robo advisor.
rvz大约 2 年前
At least Bloomberg is not pretending to be &#x27;open&#x27; or &#x27;open source&#x27; with their GPT model despite the same excuses. Unlike Microsoft® AI.com.
m3kw9大约 2 年前
Link to use model?
opisthenar84大约 2 年前
I wonder how well this performs over standard GPT-4 with domain knowledge injected in as prompt.
评论 #35382357 未加载
steve1977大约 2 年前
As quant finance has been confidently wrong for decades, that’s probably a great match!
OrdoAbChao大约 2 年前
What is the average P&#x2F;e ratio of the stocks listed in the NASDAQ.
eismcc大约 2 年前
I’ve been working on <a href="http:&#x2F;&#x2F;www.dollarsign.ai" rel="nofollow">http:&#x2F;&#x2F;www.dollarsign.ai</a> and it would be interesting to have this directly integrated.
dongobread大约 2 年前
They&#x27;ll release training logs but no model file or data? How does anyone trust that this isn&#x27;t just overfitted garbage trained directly on their benchmarks?
meghan_rain大约 2 年前
&gt; BloombergGPT far outclasses the models we evaluated ourselves, and is slightly behind GPT-3.<p>tldr it&#x27;s worse than GPT3?
评论 #35382925 未加载