CHAT GPT FOR STOCK PICKERS? Throughout the course of my 15 years as a stock-picker I've always erred on the side of being a luddite. To me, developing a great stock idea is an artisanal process and the same way a painter might find joy from cleaning his brushes,
I found joy from printing a 10-K, closing my office door, and spending an hour with a pen & highlighter. I've generally been skeptical of technology efficiency tools, and default recommend buy-side analysts build their own models from scratch & own the research process from
step 1 to step 60. After all, you never know which one-line of the model will make or break an idea, and I don't want that line distorted because I downloaded an excel file of financials w/ errors instead of keying in the numbers on my own. Smart people disagree about this.
It was kind of ironic, then, when I was hired as a consultant to help a quant fund integrate fundamental outputs into quantitative investment processes. To be honest, I didn't know the first think about machine learning (and still don't know much).
But for much of the last 5 years when people asked me "will machines put human stock pickers out of business", my general answer was this: "While alpha pools are inexorably shrinking as quants get smarter, equities are long-duration, forward looking securities and modern alpha
is generated not by observing the present fundamentals but by anticipating the future state fundamentals with a risk/reward overlay...this anticipatory & asymmetry game is not the realm of quant models, which rely on historically recurring patterns".
I still think that's mostly true. Like many of you, I've been going down the ChatGPT rabbit hole, mostly by talking to people much smarter than me. What I hear, and what makes sense to me, is that we are still years off from ChatGPT providing expert level conclusions.
If you've asked ChatGPT to pitch you a stock (and know anything about stock selection), you know that you get back mostly high level gibberish like you would see in a really bad sell-side initiation about "competitive advantage" and "solid business model", like this treat:
Did you see the many demonstrably incorrect assertions? While this might look impressive to a stock picking neophyte, I can ASSURE you this type of analysis is almost completely worthless. So, why am I writing about ChatGPT? One movement that has gained steam on the buy-side
is process efficiency. Canalyst, Tegus & AlphaSense are pioneers in accelerating the analyst flow. 10 years ago, to get up to speed on a name I'm doing 15 expert network calls and building a model. Today, I could read a stack of expert network transcripts and pull a Canalyst
model and save 10-15 hours in the "getting up to speed process on a name". This process efficiency might seem silly, but I assure you is VERY valuable. To me, this analyst workflow element is the most interesting frontier of AI.
AI NOT as a predictive tool to "pitch me a stock", but a process tool that can take my 60 hour research process down to 20 hours, providing the same fundamental value as expert transcripts & pre-built models: helping me do my job more quickly.
Rather than artisanal, by hand approach, can a "research co-pilot" tool help conduct many of these steps?
Could we pipe in financial data APIs (which seem to be growing in availability quickly) to conduct these analyses? For example, rather than cracking open the proxy statement can my co-pilot analyze mgmt comp plan and give me a page and a half of CEO incentive alignment?
Effectively could this co-pilot create what a sell-side initiation should be? Can a series of prompts train the machine? The output is an exhaustive, deep dive analysis into the company and the various steps that are common of a institutional grade buy-side research process.
The mental model to me seems to be GitHub's Co-Pilot, a "human in the loop" program where the human user can tweak or discard what is generated. It's an efficiency tool, ultimately. The same way that reading expert transcripts is an efficiency (and cost savings) tool.
Can the machine run our "baselining" exercise where we listen to the last 12 transcripts and look for deviation in tone? Can we train machines on identifying the alpha signals we look for?
Can we train AI to look "good business", "good management", "accelerating business momentum"? Could we build a much better version of Morgan Stanley's "what's in the stock" analysis overlaying sentiment, consensus & idiosyncratic performance since last catalyst?
Could we train the machine to estimate the "buy-side whisper" by integrating sentiment, volume, ownership & idio perf data? Could we have the AI take the first cut on constructing a risk/reward? Then we can accept or reject those constructs? I think it's possible.
The luddite in me never really thought this was possible. Am I'm still not sure. But call me intrigued. I do think there is a very real possibility we can cut down a typical 60-hour deep dive into 20-40 hours.
That time savings for the busy investment professional is a HUGE deal. I'm excited enough about that opportunity that I am conducting a feasibility with the mental model of code augmentation. A Git-Hub Co-Pilot for the stock picker, mimicking the general analyst flow.
Effectively the plan is to train a vector database supported AI on the Fundamental Edge Analyst process to help the human analyst get up to speed MUCH more quickly on a name. I'm working with two partners on this.
If you are geeking out on this like I am, I'd love to chat, get advice/feedback. I have a strong handle on the buy-side analyst work-flow (it's kind of my life's work at this point), but I am very much NOT an AI expert. If you are working on a similar problem, I'd love to chat.
Also, to the extent the feasibility study and prototype are promising, a co-development partnership with a large asset manager (who has existing APIs and internal data repository) seems like the right path. If that is you, please reach out via DM or brett@fundamentedge.com