Large language models (LLMs) have finally broken into the mainstream. LLM-driven tools are transformational. They’ve revolutionized consumer products (auto-suggest, Google Translate, Grammarly) and now LLMs are coming for the investment research industry.
AI is an extremely powerful and effective tool. However, there’s no denying that we’re in the midst of a hype cycle that is bordering on the absurd.
We made this post to help our friends, partners and users cut through the noise, and focus on what matters. We’ve included a list of some of the AI tools on the market with notes on their strengths and limitations.
A few general notes on generative AI
Anyone who’s used ChatGPT, BingChat etc for financial research will know that hallucination issues are rampant and get worse for technical topics. (These tools frequently invent plausible sounding answers that have no grounding in reality.) Generative AI hallucinations are hard to fix and aren’t going away tomorrow. Be skeptical of generalized generative chat bots that claim to provide factual information.
About us
Bedrock AI is a language model-driven investment research software platform. Our core language models are highly specialized (they are trained specifically to read/understand complex financial disclosure) and are not generative. Our models specialise in computational “understanding” instead of generating text. For context, our models were trained on 65X more SEC filings than BloombergGPT.
Our flagship product, Bedrock Forensic Intelligence (BFI), uncovers high-impact red flags in SEC filings and provides advanced SEC filing navigation tools. All results are available in real-time.
Learn about BFI’s core features here, our in-house NLP-breakthroughs here and our upcoming LLM-driven features including a financial Q&A assistant here.
Interesting finance-specific tools using generative AI
Hila provides question answering for specific documents e.g. for a specific earnings transcript or 10-K annual report. They cover a fixed number of companies so you don’t have to upload the document yourself. Their performance over earnings transcripts is substantially better than the annual filings, because of the smaller length of the documents.
What we like:
The context window is smaller (1 document) which limits hallucination.
They (mostly) give an error message when they cannot locate the answer in the document, instead of making up a plausible incorrect answer.
They’re up front about what they can and cannot do well. They disclose that Q&A works much better for table and numeric retrieval rather than qualitative question answering.
The cons:
You have to know which document likely contains the answer you’re looking for, which limits usefulness.
No small cap coverage.
You can upload PDFs of filings or earnings transcripts and ask questions about it like ‘List all risk factors in bulleted form’ or ‘give a summary of all legal issues facing this company’. Docalysis provides the page numbers from which the output was extracted plus their library contains some 10-Ks pre-loaded.
Unfortunately, some of the answers are still factually incorrect and not in fact sourced from the document in question.
FinChat.io from Stratesphere.io
A fellow Toronto fintech, FinChat launched a few weeks ago and created a lot of excitement on the interwebs. FinChat uses a combination of ChatGPT and langchain/llamaindex, similar to the tools above but is building for a much broader use case.
What we like:
FinChat is a great Q&A tool for pulling up financial information. FinChat pulls financial data and metrics from Strastephere.io and therefore, has more complete and accurate answers in this domain than the tools listed above.
The founders are cool ;) and seem to be iterating quickly.
The cons:
Qualitative question answering suffers from error and hallucination. For instance, we asked about server useful lives at Alphabet and got back an answer about Greenpeace.
ChatBiz.ai is a financial chatbot, as the name implies. We haven’t tried it yet. You can test it here. Once you do, let us know what you think.
Other awesome tools using NLP and AI
Babbl.dev uses language models to detect bullish vs. bearish sentiment indicators across stocks, topics, and market-wide.
We’ve been known to call other financial sentiment tools ‘purveyors of useless bullshit’ but we love Babbl. Babbl’s founders are computational linguists and computer scientists. They aren’t pretending to do lie detection (phew) and generally abstain from the insane claims that we’ve seen from other sentiment tools. We love their newsletter. It’s pure fun.
280First is the granddaddy of financial NLP tools, having been around since 2016. It is a financial services platform that analyzes unstructured data from various sources and produces insights from financial documents for investment professionals. There may be a human-review step in their process but we’re not sure about this.
Nosible AI makes it easy to demonstrate what your portfolio looks like and how it’s different/unique. They have visualization and data libraries.
BlueFire AI combines financial statement analysis, behavioural profiling, market data analytics and NLU to provide signals of underperformance in the listed assets of public companies. They also use data visualisation to contextualize information.
Roic.ai provides ten years of financial statement data in one, easy to use screen. No clue how they use AI but, let’s be real, who cares.
New Constructs uses forensic accounting and technology to provide research, tracking and alerts for an unlimited number of tickers across 50 portfolios as well as screening and direct access to their databases. They also offer API & Excel Add-Ins.
Kailash Concepts seeks to exploit pricing errors that arise from herding, confirmation bias, and self-attribution bias among others. They offer a monthly research report as well as a free “quick takes” newsletter.
AlphaSense uses machine learning and NLP to offer finance-specific smart search as well as a number of linguistic and sentiment scoring offerings.
Sentieo (owned by AlphaSense) provides a searchable database of public and private company data, documents, and news from relevant sources, along with integrated modeling tools.
Amenity Analytics provides an NLP platform that can be customized to extract and classify the information most relevant to you, capturing key business events, sentiment, context, and temporal relevance consistent with your research strategy.
Boosted.ai uses finance-specific machine learning algorithms to identify patterns based on your unique inputs. They also highlight which features drove the machine learning decisions and outcomes. Another Toronto fintech!
Bloomberg, FactSet, CapIQ, Refinitiv: All of the major capital market research platforms use AI, ML and NLP, to some extent, in order to do information extraction. FactSet and Bloomberg also provide various sentiment and linguistic indicators.These tools use a human-in-the-loop methodology for information extraction, which improves the accuracy and reliability. Bloomberg recently published a paper on their generative LLM, “BloombergGPT”. Learn about how our models compare to BloombergGPT here (towards the bottom of the article).
What did we miss? Know a great tool, tell us about it via Twitter DM.
For more of our favourite tools, visit the Tools & Resources section of our blog.