AI on the Buy Side, Part 1: Idea Generation

This is a four-post series I'll be publishing for anyone on the buy side of public equities researching AI solutions for their workflows.

We've worked with hundreds of buy side clients over the past 4 years, I'd like to share some of what we've learned, the current landscape for AI workflows on the buy side and where I think things are headed.

The four parts are:

Idea generation (this one)
Continued vetting, i.e. research
Thesis tracking
The future: where is this headed Today I'll be writing about what has been my favorite part of the research process to build for: idea generation.

Job to be done

"It's frustrating to think that there are likely dozens of great ideas out there which squarely fit into our investment strike zone …and I'm totally unaware of them"

— CIO at multi-billion dollar L/S Tiger Cub

"Screening is great, but it's just a starting point. I want to see the context, see the deep dive before assigning it to one of my analysts."

— PM at multi-billion dollar Long Only

"My holy grail? I'd love for AI to tell me where I'm winning consistently, and what trading setups I'm bleeding out on."

— CIO of $500m L/S fund

Screening has two purposes: the first is to rule out companies that are uninvestible according to your framework, and the second is to rank ideas by their likelihood of success.

It is most applicable for generalists who cover a wide universe that would be impossible for a single individual to keep track of.

Every fund or pod has a limited amount of capital to allocate, so the question is never "should we invest in X," it is "should we invest in X instead of Y or Z."

Analysts and PMs don't just need good investments in their universe, they need the best investments in their universe.

Before LLMs this meant that analysts needed to guess when their continued effort to screen companies for qualitative measures began to diminish their total return on time.

Now we have tooling that can perform complex qualitative research and rank thousands of companies in an evening. It has become a question of how to best compound research effort with inference.

Solutions

There are a few different flavors of screening that exist today for you to try or build in house.

Pure quantitative screening

The first is of course pure quantitative screening. I won't comment on the ergonomics of different providers, i.e. their query languages or GUIs, but any frontier LLM should be able to get you most of the way there when it comes to translating natural language into something the provider can execute deterministically.

Bloomberg will have the widest breadth; contenders are the major financial providers, Visible Alpha, S&P, FactSet. Ideally you use many together with fallbacks to widen your universe as much as possible.

Quantitative screens are the building block for the next two more sophisticated systems I'll describe and are almost always used as a preliminary.

Simple qualitative screening

The next flavor of screening is what I'll call simple qualitative screening. For a given filtered universe of companies you run a keyword or fuzzy search, then for each retrieved chunk of textual information you classify it with an LLM or other type of ML model according to the original input.

There's a lot of knobs and levers to turn to get this type of screening to work well — it's not as simple as it may seem. Things like model choice, search system choice, data cleaning and curation all matter.

An example might be "find me CEO changes in the last 6 months": 1,000 chunks of text mentioning CEO changes are surfaced, and each is individually classified as either yes, there was a CEO change in the time span, or no, there was not.

This method works well for execution on quick market-moving changes (e.g. are our holdings exposed to tariffs on Vietnam?) but falls short when it comes to deeper understanding and the ranking of ideas.

There's a sub-type of this type of screen where pre-computed tags, e.g. CEO_CHANGE, are applied to companies so that you can quickly filter; you lose the flexibility but gain time to response.

Companies like AlphaSense, Hudson Labs, and of course Portrait Research support this type of screening: Build with AI Assistant

These types of screens can get expensive quickly depending on your needed depth, model strength and context.

Our Solution

The most sophisticated form of idea generation is one that we're very excited about: we call it Mosaic, our agent swarm product.

Put simply, we launch an agent for every given company inside a quantitatively filtered universe, we allow agents to share research and evaluate comps, and finally we rank the results from the archive of research produced.

The benefit of this method is that the generated archive of research can be updated flexibly.

For example, if you're looking for good shorts that have revenue heavily concentrated in a single customer but all your results are benefiting from the AI data center build out, simply ask the swarm to filter out AI CapEx suppliers.

These types of screens have the highest variable cost and require the greatest effort to encode your fund's investment frameworks.

I'm not aware of any other solutions besides ours that operate this way. Please reach out if you're familiar with one, I'd love to compare and contrast.

Implications

The way I see it an analyst's job is evolving similarly to how software engineering has evolved over the past year or two. Instead of implementing code by hand we now focus on how to manage LLMs in a way to verify our deliverables are secure, ergonomic for the end user and stable. Following this analogy, analysts will in the future spend more time ensuring the system — or in this case screening framework — mines and synthesizes information they can make predictions on than the research itself.

With LLMs as the main research agent we can back test our qualitative frameworks. As an example, if our screening framework is "under performance of a business against peers along with CEO changes in the last 6-12 months, then rank," will the top half of results outperform the bottom half?

If we can show some level of predictive capability above noise, we can begin tweaking our framework and research to find out which qualitative aspects of a company change our predictive performance and how they hold up across time periods.

This all depends on having frozen data sets across a few time periods as well as frontier models with verifiable memory cutoffs.

Along with back testing, the ability to scale frameworks across a wider universe of companies means there will be more downstream research to validate and build upon. That means the tooling underneath idea generation also needs an AI overhaul. More on that in my next post.

Thanks for reading! Feel free to reach out if you have thoughts or you'd like to connect.