Why AI Won't Replace Stock Pickers - Jack Woodruff, Candlestick & Vinoo Ganesh, Kepler

A fundamental investor can understand a company thoroughly and still have no reason to trade its stock. What would the investor expect the business to do differently from what the market expects, and why? Jack Woodruff, founder and former chief investment officer of Candlestick Capital Management, calls that disagreement a divergent view. In conversation with Kepler CEO Vinoo Ganesh, he expects AI to change how much information every investor starts with. Investors still have to find a disagreement worth acting on.

At Citadel, Woodruff had 18 people covering 400 stocks. At Candlestick he had 30 stock pickers. Could one great analyst cover all 400 by 2030? Consider the other extreme first. Spend all your time on one company and you might know it better than anyone, only to discover there’s no trade to make. Split your time evenly between two and each gets half as much attention. Software can prepare material on many more companies, but the person making the decision still has to absorb it.

Woodruff imagines AI doing much of his usual six to eight hours of preparation in five minutes. What happens to the hours saved depends on the competition. One analyst might finish the old workload in two hours. Another might still put in a 16-hour day, using the extra time for channel checks and calls with people who know the industry. The second analyst has a chance to learn something the first one never asked about.

Ganesh asks which part of a junior analyst’s week would look absurd in five years. Woodruff would keep reading on the calendar. How much work an analyst can hand off depends on what they need to learn by doing it.

• Following a company closely. Summaries can shorten the time spent reading 10-K annual filings and earnings-call transcripts. The analyst who follows the company still needs to understand the material well enough to weigh the next piece of information.
• Working through large datasets. Finding, cleaning and checking data can take up much of the week. Woodruff sees more room to delegate that work. An agent could also flag companies that warrant a closer look, giving the analyst a reason to open one filing before another.

Once every fund can get the same cleaned data, having it becomes the price of competing. The time saved can go into another proprietary research call or into building a fuller picture of an industry. Woodruff’s investment mosaic typically has three or four observations that point in the same direction and can come from different sources. One reason to make the call is to find out whether a pattern in the data also turns up elsewhere.

Woodruff started his career as a CPA. Would he sign off on a model’s number without seeing where it came from? His experience with human data entry explains his refusal. Errors were rare, but one bad entry could cost millions of dollars if a fund acted on it. He would need to trust the figures before he would let an AI system use them to direct trades.

Investors Woodruff knows ask Claude to help build financial models. Once a model arrives, someone still has to go through the numbers, check that they’re all there and establish where they came from. That work belongs in any estimate of time saved. The feature Woodruff values at Kepler is the ability to audit each figure.

Ganesh describes Kepler’s implementation separately. The language model interprets the question. Deterministic code then retrieves the source data and performs the calculation, so an analyst can trace a figure to its inputs and reproduce it. Introducing Kepler

A buyer has a second question for an AI vendor. Where will the fund’s internal data go? Woodruff would leave the technical decision to his CTO. Ganesh gives the requirement a name, data isolation. Woodruff wants the fund to keep control of the research it did for itself.

Getting the data together and automating routine preparation is the first stage Woodruff and Ganesh envision. A fund can produce an earnings preview or a discounted cash flow valuation and still have no reason to take a position. Once those materials are readily available, a more valuable question for the software is which company deserves another day of an analyst’s time. Even an opportunity the fund keeps to itself can be short-lived. When other investors see someone making money, they have an incentive to work out how.

Woodruff’s old team did an enormous amount of preliminary work whose most promising result was a research assignment. Something about one company looked interesting, few people seemed to be investigating it, and a particular outcome could be worth a great deal to the fund. The analyst now had a focused question for fieldwork. How much reason was there to believe that outcome would happen? Even with the same models available across firms, investors can ask different questions, seek out different evidence and reach conviction at different times. An analyst who spots the question early has more time to investigate it before competitors arrive at the same view.

Asked what he would build and what he would buy, Woodruff makes one easy purchase in front of Kepler’s CEO. He’d buy Kepler. The harder decision is who to hire. His old firm had a pyramid. Many analysts presented research to senior stock pickers, who in turn presented it to a couple of portfolio managers, with the chief investment officer at the top. Preparing information for the next person accounted for much of the work at the base.

Senior stock pickers would still have a place on his team. The open question is what expertise to put alongside them. Woodruff can imagine hiring investigative journalists for deep research projects or data engineers to work with the information the fund collects. He might need fewer people to assemble standard materials. He can also see a reason to keep analysts. If software makes common information cheap to process, there may be considerably more work for people who can obtain information the fund can call its own.

Who benefits more from AI, a five-person fund that can act like 50 people or a 5,000-person firm that can act like five? The mandate matters. A fund that holds a company for five years has a different appetite for new ideas from one that trades throughout the day. Woodruff imagines the five-person team running a concentrated portfolio because it can thoroughly investigate only so many companies.

What changes if the team can research more companies while keeping 20 places in its portfolio?

At 400 stocks, only one company in 20 has to earn a place. The team can pass on more marginal ideas. Whether that improves its hit rate depends on the research behind each choice. If the team knew the original 100 companies well and has only skimmed the other 300, the smaller selection percentage says little about how often its investments will work.

A 5,000-person firm might run several strategies across different asset classes. Its advantage is harder to express as a single coverage ratio. What mattered to Woodruff was idea velocity, the rate at which good, researched ideas became available. When a position worked and the fund took it out of the portfolio, it freed capital for another investment. Having the next idea ready was part of how the team sought to earn returns over time.

Woodruff can imagine a collection of AI agents replicating the work of a fundamental fund that existed a year ago. His objection concerns which funds it will have to outperform next year. Human-led funds get the same preparation tools. While the agents reproduce yesterday’s research process, competing stock pickers can spend the hours they recover investigating the next question.

Quantitative funds have traded without a human making each decision for decades. Woodruff’s prediction concerns fundamental investing, where someone has to decide which business question warrants original research. If agents could take over that judgment completely, he expects the contest would come down to who had the most computing power and information. His own hiring bet is still on the person who knows which company is worth the next call.