Kalshi and the Search for an Orthogonal Signal

Prediction markets have spent the past two years moving from the margins of financial markets towards something institutional desks are willing to take seriously. Kalshi, the CFTC-regulated event-contract venue, has been at the centre of that shift, and its ambitions now extend well beyond providing a place to trade. Over an eight-week period the firm has stood up a data business, distributing its pricing through Tradeweb, out to Bloomberg, and – subject to the technical work still in progress – through LSEG.

Why might the signal be uncorrelated?

The argument for prediction-market data as an alternative source of alpha rests on incentives. A sell-side economist who publishes a forecast a long way from consensus carries a real professional cost if the call is visible and wrong, and often a cost simply for standing apart from peers. That pressure pulls published forecasts towards each other. A price set by anonymous participants with capital at stake removes the reputational penalty and rewards accuracy alone, which is what should give the resulting signal its distance from conventional forecasting.

Andy Ross, Head of Institutional at Kalshi, describes a market that corrects itself when weight of money pushes against informed opinion. “Immediately after an event – a leadership debate, say – people come in and try to buy a candidate to drive their odds, because that candidate appeared to do well,” he tells Market & Alt Data Insight. “The odds move, but within about 20 minutes everyone else pushes the price back to fair value on the wisdom of the crowd. For a short period you can move the market by weight of money, but the market price reflects the accuracy of the consensus of the smart people in it, because you get paid for being right.”

What is the data actually telling you?

Part of the appeal for a data buyer is the ability to price a single variable in isolation. A share price bundles a great deal of information – past sales, forward guidance, management quality, the probability of a takeover – into one number, whereas a contract on a discrete outcome prices just one of those elements. Ross provides an example.

“Home Depot had record sales, record revenue, record profitability – and when they released them, the share price fell by 5%,” he says. “If you can trade on sales directly – the store’s always busy when I go in, so I’ll buy sales – you’re not buying the stock price, you’re not buying whether the CEO had an affair, you’re not buying whether there’s an accounting write-down. You’re just buying sales. You can atomise a stock into many different risk elements.”

Used this way, the data becomes an input to valuation rather than a hedging instrument, and its attraction is diversification: a clean, single-variable estimate that sits alongside the proxies – card spend, footfall, web traffic – a desk already runs. Ross says a small number of firms are already consuming Kalshi markets in this manner, naming Ark Invest as a public example.

The firm also makes a strong claim about the quality of that data. Ross points to Brier scores – a standard measure of probabilistic forecast accuracy, on which a lower number is better – putting Kalshi’s markets at a Brier score of 0.05 or below a month out once somewhere between $40,000 and $60,000 has traded on a contract, against about 0.1 for what he describes as the best meteorological models. On this reading, that makes the market’s forecasts nearly twice as accurate as models the world has spent heavily to build.

It is worth stating that this is Kalshi’s own measurement rather than an independent audit; Ross explicitly sets aside election and sports markets; and it applies to well-traded, calibrated contracts, which are a subset of the roughly 8,000 markets the firm reports running rather than the book as a whole. Within those limits – thickly traded contracts on measurable outcomes – the calibration claim seems a reasonable one. Read as a statement that prediction markets outperform established forecasting in general, it goes further than the underlying figures. The Federal Reserve’s recent working paper on macro markets (FEDS 2026-010) sits in similar territory, finding prediction-market pricing competitive with established forecasts on particular measures rather than superior across the board.

Does the signal survive institutional flow?

An alternative dataset derives much of its value from being uncrowded. As more capital enters a market and pricing becomes more efficient, the edge available to any one participant should narrow, which raises the question of whether institutional adoption erodes the very signal that attracts it. Ross does not dismiss the concern.

“Could we get to the point where it’s leveraged and 50, 100 or 1,000 times bigger, and the data quality goes down? Maybe – we’re not immune to that,” he says. “But fundamentally, you’re still paid for being right. What you have is people intensely focused on being right, and if more people are wrong, the right ones simply push harder against them.”

A couple of things to note here. As a trade, the edge may well compress: Ross notes that hedge funds are already arbitraging the gap between Kalshi’s inflation and payroll pricing and the rates implied in the TIPS market, and spreads of that kind narrow as more money chases them. As data, the same deepening of liquidity could make the market-implied probability more reliable rather than less, provided the correction mechanism he describes continues to hold. A desk buying the data to sharpen a valuation and a desk trading the contracts for alpha are therefore asking different questions of the same market.

Can a desk actually consume it?

None of this is usable if the data cannot be sourced in a form a quantitative team can work with, and here Ross is candid that the business is early.

“We’ve stood up our data business in eight weeks, so this is a moving target,” he says. “We’re not yet sharing level-three data – we just haven’t built it, though there’s palpable demand for it, to train models and study liquidity. Our aim is to have the data widely available. Right now I’m not interested in monetising it – I want to maximise the number of eyeballs on the data, whether human or machine.”

Real-time level-one and lightweight level-two data are available today through an API and a WebSockets feed under a data-only licence. Full order-book depth is not yet a product. More significantly for anyone intending to back-test a strategy, there is no clean historical delivery through the kind of cloud infrastructure institutional desks expect – access through Snowflake or similar is described as a roadmap item rather than a current capability, and historical data is supplied on an ad hoc basis. For an investment team, a signal that cannot be tested against a point-in-time, survivorship-complete record is difficult to trust in production, however well the live feed performs.

The regulated structure does give the eventual product something offshore venues cannot easily match on the question of provenance. Kalshi operates as a CFTC-designated contract market and clearing organisation; every participant passes through AML and KYC checks that tie each trade to an identified counterparty; and settlement on its crypto contracts uses an independent, trade-weighted benchmark designed to resist manipulation. For data intended for use inside a regulated investment process, that auditability obviously matters.

Whether the signal remains uncorrelated as the institutional flow that would make it worth consuming arrives is a question that cannot yet be answered, because the historical record needed to test it does not yet exist. Until it does, the buy side is left to weigh a plausible mechanism against a data business that is, on its own account, still only eight weeks old.