Kirk McKeown

The through-line

Everything Kirk McKeown says circles back to one claim: alpha is competitive advantage, and competitive advantage is not a fixed thing you own but a moving target you have to keep re-manufacturing. "Alpha in 2013 is different than alpha today. Alpha in 2006 is different than alpha in 2013. Alpha moves around" [2]. Sometimes it sits in speed to information, sometimes in access to the information itself, sometimes in an organisation's ability to process something and turn it into a trade [2]. Edge is rarely static and often lives in process design, information capture, and the interpretation of small narrative inflections [3]. The corollary he keeps returning to is that competition erodes it: "competition lowers competitive advantage because more people in um the arbitrage and spreads get tighter" [2].

The second half of the through-line is what follows from the first. If edge migrates toward differentiated inputs and the synthesis of them [3], then the binding constraint on most investors is that they cannot get enough inputs. That belief, formed across twenty years of building research functions, is the direct origin of Carbon Arc: "my whole belief was if you could build a portfolio of inputs, you're going to have better outcomes over time. And the only way you could do that is by breaking price" [1].

On the three levers every research input has to move

McKeown reduces the evaluation of any PM, and therefore of any research product or dataset sold to one, to three variables: "Number of at bats, hit rate against set at bats, and then sizing against that hit rate" [2]. Any input entering the conversation has to create lift in one of them, or it is worthless. "If you're not creating lift in one of those three buckets for a for a hedge fund when you're running a research business, you don't have a research business" [2]. The same arithmetic governs how a fund should price data: a hit rate moving from 55% to 57%, a hundred at-bats becoming 115, or a million-dollar bet becoming a million and a half on the names where the data creates conviction, is where the lift comes from [1].

Of the three, he prefers to play in the middle. Hit rate is the lever he finds most tractable and, crucially, the one you can measure cleanly [2]. From this he builds a valuation framework for data: price against "the number of informed tickers, the number of informed at-bats, the number of those at-bats that worked," done consistently enough over time to construct "almost like an alpha cash flow framework" and back into what an asset is actually worth [1]. He is explicit that the answer differs by strategy, and that his framing applies most naturally to funds trading on horizons of roughly fifteen days to six months [1].

On running research without a P&L

McKeown is emphatic about the separation between a research function and the investment book it serves, and he defends it on grounds of intellectual hygiene rather than compliance alone. "You want to be really focused on getting it right. And you get it right through process. You don't want to be right. You want to get it right" [2]. A research group that starts thinking of itself as inside the P&L framework invites confirmation bias, and the separation of church and state is what "keeps the research clean" [2].

That separation also means accepting you will never get clean attribution. Whether your work was 5% or 100% of a decision is unknowable, and that is fine [2]. What is not fine is scoring yourself on the wrong metric. "You can't say hey the stock was up 10%. That's actually not the right metric" [2]. If you called out that a launch was slower than expected and it was, that is a win even if the stock never moves. "You're not getting paid on the return. You're not getting paid on the atbat. You're getting paid on the Grinch" [2]. He couples this with a demand for brutal self-assessment: track it maniacally, be honest about right and wrong, because "you scoring yourself hard is better than you know better than anything else" [2]. The researcher's competitive advantage, in his account, is genuinely different from the PM's: research goes deep on value chains and supply chains, while PMs and analysts have to put everything together, which he considers the more nuanced and difficult job [2].

On what different firms optimise for

McKeown reads each firm he worked in as an alpha-generation machine tuned to a particular driver. At Tudor Investments during the internet cycle, with far fewer firms competing, edge came from domain expertise and analysts who stayed close to their names and built good cash flows [2]. At Glenview Capital, a two-year-horizon shop with roughly half of committed capital in the top ten positions, you had to know your names cold and understand the thematic drivers and unit economics behind them; the modelling culture was deep, and the firm's reputation as a value-added partner to management teams was itself an input [2]. He calls that "a slugging game": fewer bets, so they have to be bigger, and sizing has to be right even though hit rate still matters [2]. Point72 he describes as the opposite optimisation: catalyst-driven, variant-view investing, turning the book far more, and "at a place like 72, it's a hit rate game. If you're taking a lot of at bats, you need to make sure your swing is really tight" [2][3].

The edge he attributes to Point72 is perceptual granularity plus organisational repeatability: "it was about understanding when stories changed one degree instead of 10," knowing where the market is, where the world is, and whether the story has moved against new information or a shift in management tone [2]. What made that scalable was a top-down insistence on "repeatable, transparent, rigorous process" across an organisation with 120 teams investing in different ways [2]. His generalisation is that alpha can be manufactured at three levels, organisational, people, and process, and sustainability comes from optimising all three at once [2]. He is direct about the calibre of operator this requires: Steve Cohen, Ken Griffin, Izzy and Dmitri are "not just picking great stocks, but building phenomenal businesses" [2].

On why the data market barely trades

The friction list McKeown gives for how funds buy data is specific and cumulative. Trialling, sampling and long sales cycles come first, and are inconsistent across the space [1]. Then valuation, working out what a dataset is worth relative to price and what it will actually do for the business [1]. Then usability, which takes specialist skills to clean, normalise, tag and dedupe raw panels [1]. Then licensing, with terms of use that consume time and some licences a firm simply will not sign [1]. Then compliance: how the data was collected and where it lives [1]. Then pricing, which is onerous when only a few players hold the market [1]. Stack those together and "data doesn't really trade, it's really expensive, it's really hard to work with, and, you know, the lead times and sales cycles are really long" [1].

He describes the resulting equilibrium as self-reinforcing and bad for both sides. Providers keep prices artificially high; consumers are forced to make large decisions on very few assets, which crowds out everything else they might have bought [1]. A fund with three to five million dollars to spend "can chew that up with three or four data sets," covering roughly a hundred tickers [1]. Meanwhile the assets that did aggregate into discrete pools, credit card, clickstream, geo and app, carry pricing that is mismatched to the alpha they actually generate, a legacy of how data grew up on Wall Street [1]. And the differentiation decays: "credit card data in 2013 was a real differentiator... Today it's table stakes," to the point that a consumer trader without it does not even know what peers are trading against [1].

On pricing insight like an option

McKeown's argument for a market-based approach to data starts from what an aggregation actually is. Take any panel, credit card, clickstream, healthcare claims, do the work of cleaning, normalising, tagging and deduping, and you produce an aggregation that has time decay, that turns out to be right or wrong, and that carries a cash flow. "Those are the three things you need to price an option. So fundamentally, you're you're writing tickets. You're it's a derivatives business" [1]. If it is a derivatives business, you can take a market-based approach to the underlying [1].

The Carbon Arc design follows from that diagnosis: smash down the cost of insight, break the licensing construct, pay providers on what gets consumed, and let consumers build the data structure they need to drive hard ROI on their buys [1]. The intended result is more consumption, more diversity of inputs, and a consumption-based framework for insight generation [1]. The underlying economics he states plainly: "Supply and demand are too far apart and the ability to pay and willingness to pay are the biggest drivers of data liquidity" [1]. His starting observation is that "for all the data that exists in the world, very little of it makes it into the hands of decision makers" [1].

On the size of the market, and where the wallet goes

McKeown sizes the alternative data market by backing into the revenues of the big providers in card, click, app and geo, excluding market data: cards perhaps a couple of hundred million dollars in financial services, app maybe another hundred, clickstream visible through a public comparable, for something on the order of 500 million across the big four [1]. The buyer count is the more striking number. He estimates roughly 500 qualified buyers on Earth for what he calls data capex above two million dollars, including big tech, large hedge funds and financial services firms [1]. His ambition is to take that from 500 to 500,000: "I'd rather 500,000 at 1,000 bucks than 500 at a million. Right? You know, it's better business. But our market's healthier" [1].

He is careful not to overclaim growth. In the near term he thinks this is largely a redistribution of wallet across Wall Street rather than a bigger TAM [1]. The variable that could change that is agents. If the analyst pool stays stable, one thing follows; if it actually shrinks, he expects some of the capital previously paid to analysts to be reinvested in data structure [1].

On portfolios of weak signals and the ontology hiding in transaction data

His research method has always been an accumulation rather than a scoop. "I never talked to one person and said, 'This is what's happening.' It was always an amalgamation of weaker signals. I'd rather five weak signals than one strong one over time" [1]. He acknowledges he could do that only because he worked at shops that believed in proprietary research and gave him scale, and that most firms cannot; making that process available to them is the point of the business [1]. The same instinct shows up in how he broke through early: locked in a computer lab, cold-calling hundreds of people around a merger, reaching supply chain heads at Scott's Miracle-Gro, Whirlpool and Procter & Gamble, and building a model off those calls that produced a synergy number of 1.5 billion against a published 500 million [1]. Notably, he built out the primary research business at Glenview partly because he was not as comfortable in Excel as colleagues who had come through banking and private equity [2].

One underrated use of data he flags is structural rather than predictive. A receipt panel contains millions of SKUs, brands and products, which makes transaction data a quasi-reference asset for building the ontologies and taxonomies a firm needs internally as it constructs a broader data capability [1]. That is a different justification for buying data than signal generation, and he thinks it is under-exploited [1].

On apprenticeship, AI, and how young analysts get trained

The thing McKeown says scares him most about proliferating AI tools is not the models but the loss of the apprenticeship. The first two years at a bank, or in an analyst academy at a large fund, are spent in Excel learning to build a three-page model with cash flow and balance sheet, and that now "happens in 9 seconds with 15 different plugins right now" and the model is clean and works [1]. He grants that automating and enriching tools are genuinely valuable, but insists "nothing beats doing the work" [2]. If you have never pulled apart a three-sheet model or spent time in the MD&A and a 10-K, the value of that hammer-swinging is lost [2]. His worry is a "scuba diving versus surfing" shift, people assuming ownership of something they have looked at for fifteen minutes [1].

He grounds this in his own economics of effort. He worked six-hour Sundays from around 2006 to 2020: fifty Sundays a year is 300 hours, six extra weeks, "so if I'm working 13 and a half months a year to somebody else's 12 months, it doesn't matter how good your process is or how smart you are, you're not gonna beat me to the ball" [2]. And the returns increase, because in a knowledge research business "knowledge compounds," shortening time to answer and time to alpha year over year [2]. That is why he insists "Wall Street is a block and tackle business and it is a time and seat business," where "being smart isn't enough. There are a lot of smart people that don't make money in markets" [1]. He is self-aware about the risk of sounding like an old man, noting that people said the same things when the calculator arrived, and that at 49 he may be spouting old-person language about AI [1][2]. Even so, he expects consequences: a reallocation of people inside funds, and pressure on the sell-side-to-buy-side pipeline that has historically selected the best fundamental investors [1].

Takeaways

• Treat alpha as a moving target, not an asset: it has lived in speed, in access, and in organisational processing at different times, and competition continuously compresses whichever one is currently working [2][3].
• Judge any research product or dataset against three levers only: number of at-bats, hit rate on those at-bats, and sizing. If it moves none of them, it is not a research business [2].
• Price data by building an alpha cash flow framework from informed tickers, informed at-bats and how many worked, rather than by list price against an aggregated asset class [1].
• Keep research separate from P&L. Score the call, not the stock move, and accept you will never get clean attribution, because that separation is what protects against confirmation bias [2].
• The data market's problem is liquidity, not supply: usability, licensing, compliance and concentrated pricing keep most data out of decision makers' hands, and a fund with five million dollars can exhaust it on three or four datasets covering about a hundred tickers [1].
• An aggregation has time decay, an eventual right-or-wrong outcome and a cash flow, which are the three inputs to pricing an option, so insight can be sold as a consumption-based, market-priced derivative [1].
• The qualified buyer pool for large data capex is roughly 500 firms worldwide; the opportunity is redistributing that wallet across 500,000 smaller buyers rather than growing total spend [1].
• Six-hour Sundays across fifty weeks buys six extra working weeks a year, and in a knowledge business that advantage compounds into faster time to answer [2].