Proving Alpha: How a Hedge Fund Measures AI's ROI - Hari Kumar, Contour Asset Management

Loading slides…

Read the talk

How a Hedge Fund Measures AI’s ROI

Hari Kumar introduces Contour’s approach to measuring AI’s impact in fundamental investing.Open full source frame
Hari Kumar introduces Contour’s approach to measuring AI’s impact in fundamental investing.

Hari Kumar explains how Contour Asset Management compares research workflows, tracks cost per useful insight and counts the time lost to broken tools while waiting for investment outcomes to emerge.

From a talk by Hari Kumar

At a glance

Ideas worth remembering

  • Document research with and without AI at the level of data sources, delegated activities, conclusions and time spent. Contour’s concrete example follows one company between two earnings reports.

  • Process improvements can become visible before investment returns. A thesis may take six months or a year to affect a stock price.

  • Agents can preserve the usefulness of earlier research by checking shelved ideas against the market and bringing them back to the analyst, portfolio manager and trader.

  • Cost per useful insight connects analyst effort, the tools used, the ideas produced and their eventual outcomes. Tool failures and learning time belong in that accounting.

The spending happens before the investment result

An AI tool can change an analyst’s work today while the investment thesis takes months to affect a stock price. That delay makes a straightforward question surprisingly difficult: what is the fund getting for its spending? Hari Kumar, head of AI at Contour Asset Management, describes the problem at a fundamental long/short hedge fund with roughly $6 billion under management and fifteen to twenty people, focused largely on technology, media and telecom.

Kumar’s mandate starts with research and extends into trading, operations, compliance and finance. The measurement problem has two parts: whether AI helps produce better ideas and better working processes, and whether those changes eventually contribute to profit and loss, or P&L. The first can become visible while the second is still developing. That distinction drives Contour’s experiments.

0:121:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Compare the research process over an earnings cycle

Contour has around ten analysts, each covering a portion of its investment universe. Kumar describes the small-fund pattern of one analyst following thirty to forty companies, with established habits for developing and testing investment theses. Introducing AI raises a counterfactual question: would the analyst have reached the same conclusion without it? Contour approaches that question through what Kumar calls live A/B tests, documenting research with and without the new tools. 3:05

One comparison follows a single company over a three-month period, from one earnings report to the next. The baseline records the analyst’s workflow without LLMs and without additional workflows such as agent-conducted consumer surveys or tools for expert interviews. The record includes which data sources the analyst finds and what conclusions follow. Contour then runs the research process with AI tools, involving the portfolio manager, or PM, and introduces tools individually as they become available.

What changes when the agent takes on another part of an expert interview? Kumar separates two arrangements:

  • Agent interview, human synthesis: An agent conducts the expert call, and the analyst synthesizes the notes.
  • Agent interview and synthesis: The agent also synthesizes the notes, leaving the analyst to draw the final conclusions.

The distinction makes the division of work inspectable. Conducting a call, condensing its contents and forming an investment judgment are separate activities; automating one does not establish the value of automating the next.

Contour’s counterfactual comparison separates investment decisions, research access and outcomes.Open full source frame
Contour’s counterfactual comparison separates investment decisions, research access and outcomes.

The practical comparison asks how long it would have taken to source the information manually and whether the resulting conclusions change. Kumar describes the approach as useful. His slide states the limit directly. “No clean control group. No easy counterfactual.” The comparison helps Contour examine research time and judgment, but it cannot establish how much AI contributed to investment returns.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:06 · section reference included

Measure useful changes before the thesis pays off

The earnings cycle gives the comparison a workable observation window, but a thesis may need six months or a year to show up in a stock price. Contour therefore looks first at the work itself: whether consumer surveys improve, whether investment models become easier to update and whether information takes less time to obtain. Those changes can be evaluated before the market resolves the thesis.

The uncomfortable alternative is that another tool simply adds “more thrash” to the analyst’s day. A workflow can generate more material while making research harder to complete. Contour’s question is whether the tool changes how someone works usefully enough to justify the extra activity.

5:055:35
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:05 · section reference included

Keep a shelved idea from missing its moment

Fundamental research starts with a thesis, but it also requires letting new evidence loosen the analyst’s attachment to that thesis. Contour uses idea-generation agents to add further considerations to ideas organized by company and analyst. An idea may begin with alternative-data observations or an industry expert’s account; the additional work examines how it might develop over the next few months, or why the market might take longer to reflect it.

A different failure happens after an idea has already received attention. An analyst works on it for a day or two, the PM finds something else more pressing, and the idea disappears into the notes. Months later, the team may realize that it missed an attractive price window. The research was preserved, but the opportunity to use it was lost.

The proposed remedy is to have agents track those notes against the market and resurface an idea when it becomes interesting again. An alert gives the original analyst, the PM and the trader a reason to revisit the work.

This expands the job beyond generating fresh ideas. A fund must choose among more possible investigations than its analysts can pursue, and a previously uninteresting idea can become useful at a different price or time. Kumar describes entry and exit timing as an area Contour is actively exploring with agents, including positions that are no longer top of mind. 8:18

Resurfacing research when the market changes

Saved analyst notesinputs — Includes ideas investigated briefly and then set aside.
Market conditionsinputs — The context against which an old idea is reconsidered.
Agent reassessmentTracks the notes against the market and asks whether the idea is interesting now.
Analystrecipients — Receives a prompt to revisit prior research.
Portfolio managerrecipients — Receives the resurfaced idea.
Traderrecipients — Receives the resurfaced idea.
Saved research and market conditions inform a reassessment. An interesting idea can then be brought to three people independently; the arrows show information flow, not trade authorization.
Read the diagram as text
  • Saved analyst notes. inputs — Includes ideas investigated briefly and then set aside.
  • Market conditions. inputs — The context against which an old idea is reconsidered.
  • Agent reassessment. Tracks the notes against the market and asks whether the idea is interesting now.
  • Analyst. recipients — Receives a prompt to revisit prior research.
  • Portfolio manager. recipients — Receives the resurfaced idea.
  • Trader. recipients — Receives the resurfaced idea.
  • Saved analyst notes → Agent reassessment: Supplies earlier research.
  • Market conditions → Agent reassessment: Supplies market context.
  • Agent reassessment → Analyst: Flags an interesting idea.
  • Agent reassessment → Portfolio manager: Flags an interesting idea.
  • Agent reassessment → Trader: Flags an interesting idea.
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:19 · section reference included

Bring expert accounts and spending data together

Funds already buy specialized datasets, including satellite mapping and granular credit-card data. The integration problem is that each can remain isolated from the rest of the research. Kumar’s concrete example is credit-card panel data supplied by Yipit: Contour feeds it into agents through MCP and API access alongside expert-call transcripts and research notes.

Combining those inputs lets the research consider how an expert’s account interacts with observed spending data. The change is from separately consulting a transcript and a dataset to bringing both into the agent’s analysis. Kumar describes that integration as implemented at Contour, without specifying a matching algorithm or a particular investment conclusion produced by it.

8:489:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:48 · section reference included

Track cost per useful insight

Subscription and AI spending capture only part of the research cost. Contour also tracks what analysts spend their days doing: time spent on each company, insights generated per day and the eventual P&L outcomes of decisions. A fundamental fund may make a few trades a day while expecting those investments to develop over months, so keeping the research history connected to later outcomes matters. 10:35

Contour calls one of its metrics “cost per useful insight.” Its supporting questions connect effort to consequence:

  • Time: How much analyst time went into researching the company and generating the idea?
  • Usefulness: Where did the idea eventually lead, and which ideas proved interesting?
  • Tools: Which tools helped produce it?

Kumar does not specify a formula, a rule for classifying usefulness or a measured value. The practical contribution is the connected record: daily research effort can later be examined alongside the ideas it produced and their outcomes.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:05 · section reference included

Give trading and operations more time for judgment

The research experiments sit within a wider effort to improve how the firm works. Trading, operations and finance can free time for better processes and contribute to the fund’s understanding of the market. Kumar describes two distinct directions:

  • Trading: Contour measures entry and exit timing and examines how positions are spread across industries and individual companies. Agents track those companies to keep traders informed. The aim is for traders to contribute ideas to analysts and associates as well as execute orders.
  • Operations and compliance: Work is distributed across Bloomberg, order-keeping systems and other provider tools. Contour is starting to integrate activities such as order reconciliation and vendor-information tracking into one tool, reducing the need to visit multiple platforms for different pieces of the order flow.

Individual providers may already offer helpful AI features, while the work of joining them remains difficult. Consolidation is a priority because repetitive handling consumes time that people in trading and operations could apply to generating ideas, understanding the market and judging timing. Kumar presents this as a direction for reallocating skills, rather than a quantified return from a completed operations rollout.

11:4412:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:44 · section reference included

Make existing subscriptions fit each analyst’s work

The next question is how to get more value from data the fund already buys. Kumar names specialist subscriptions such as AlphaSense and Yipit alongside broader platforms such as Rogo and Hebbia. In his AlphaSense example, an analyst traditionally visits the provider’s portal, retrieves information and uses it in an existing investment model. He describes Rogo as taking integration further by bringing data into the model and providing agents the analyst can chat with.

Contour’s in-house ambition keeps the underlying subscriptions and brings their data into an internal platform tailored to each analyst’s habits and model patterns. MCP and API connections provide access; the intended additional value comes from fitting that access to how a particular analyst works. The platform is meant to learn over time from data use and investment-model updates, though Kumar does not describe the learning mechanism. This is a customization goal, not a demonstrated model-training result.

14:3515:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:35 · section reference included

Count the time spent fixing and understanding the tools

An internal platform also brings an operating cost that is easy to leave out of the ROI calculation. Model calls fail. A connector can break every couple of days. Research pauses while someone fixes the problem, and some bugs require help beyond what the technical team can resolve alone. Contour counts those interruptions and the time taken away from other work. 16:40

Learning how the tools work also takes analyst time. That cost belongs beside the time used to generate a valuable insight: a faster research step can still carry setup, explanation and repair work elsewhere. Kumar says the time spent understanding the tools is decreasing as Contour proceeds, but it remains something the firm measures.

Kumar closes with a bet on a unified layer serving multiple analysts and tasks, especially at smaller fundamental funds. Its value depends on making the firm’s existing data and people more effective while accounting for the effort required to keep that layer useful. The maintenance burden is part of the investment.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:40 · section reference included

Resources

Transcript download
Full transcript

Loading transcript…