←BACK TO BLOG
>

The Cheapest Edge Is a Market Disagreeing With Itself

Prediction MarketsMarket MicrostructurePolymarket
JUL 2026·6 MIN READ

#TWO WAYS TO BE EFFICIENT

When people ask whether a prediction market "works," they usually mean one thing: did it call the outcome correctly. That's the wrong question to lead with, because getting the future right is partly luck and entirely unfalsifiable in real time; you only find out after the fact, once, per event. There's a second notion of efficiency that has nothing to do with clairvoyance and everything to do with arithmetic: do a market's own related contracts agree with each other, right now, before anyone knows how the story ends.

The 2024 U.S. election is the cleanest illustration of why this distinction matters. Polymarket, the highest-volume and most-watched market of the cycle, correctly called the winning side in only about 67% of its markets, versus 78% for Kalshi and 93% for the much smaller PredictIt. Volume did not buy accuracy. Worse, researchers documented moments where Kalshi and Polymarket showed materially different probabilities for the identical outcome at the identical time, one platform favoring Trump, the other sitting on Harris or a toss-up. That's not a forecasting miss. That's two supposedly efficient markets failing basic internal consistency, in public, with billions of dollars watching.

#THE NBA AS A CONTROLLED EXPERIMENT

Election night gives you one noisy data point. If you want to actually measure internal consistency, you want a market with hundreds of independent trials, second-by-second data, and contracts that are mechanically linked rather than loosely correlated. That's what a UCLA paper (Cheng, Yang, Zou; arXiv:2605.00864, April 2026) did with Polymarket's NBA markets: they reconstructed continuous order-book states from more than 75 million limit-order snapshots across 173 games, then went hunting for two distinct kinds of arbitrage: the kind where a single contract disagrees with itself, and the kind where a set of related contracts on the same game don't add up.

This is a good venue for the question, and not just because NBA volume on Polymarket happens to be real money now, roughly $51 million in 2024 growing to about $0.89 billion in 2025, close to 30% of all sports activity on the platform. It's a good venue because live sports markets update every few seconds for two and a half hours, hundreds of times a season, which is exactly the repetition you need to tell the difference between a fluke and a pattern.

#THE RECEIPTS

The split between the two failure modes was stark.

  • ▸Single-market arbitrage (a contract disagreeing with itself): 7 executable episodes across the entire 173-game sample, closed in a median of 3.6 seconds.
  • ▸Combinatorial arbitrage (related contracts on the same game disagreeing with each other): 290 active episodes, with a median executable return of 101 basis points, concentrated overwhelmingly in the final minutes of live play.
  • ▸Execution reality: 76.9% of those combinatorial windows were bottlenecked by thin order-book depth, executable at an average size of only about 14.8 shares.
  • ▸Scale of the underlying market: NBA volume growing from ~$51M in 2024 to ~$0.89B in 2025, inside a Polymarket total that went from ~$73M (2023) to ~$9B (2024, with a $2.63B single month around the election).

Same-Contract vs. Cross-Contract Mispricing

7Single-market arbitrage290Combinatorial arbitrage

Combinatorial arbitrage across related NBA contracts turned up 290 times (about 41x more often than single-market mispricing's 7 episodes) and paid a median 101 basis points, while the rarer single-market anomaly vanished in a median of just 3.6 seconds, showing the durable edge lives in cross-contract relationships, not any one price.

#WHY 3.6 SECONDS IS THE HEADLINE NUMBER

Seven episodes in 173 games, gone in a median of 3.6 seconds, is basically a market policing itself. Closing that kind of gap takes zero forecasting skill; you don't need a view on whether the Celtics cover, you just need to notice a contract priced at 62 cents implying something a mechanically identical version of it prices at 65, and trade the difference. It's the cheapest edge in finance: not being smart, just being fast and paying attention. The fact that almost nobody gets to collect it tells you something a Twitter thread about "prediction markets are efficient now" would miss: it tells you bots and market makers are watching Polymarket's NBA order book closely enough to kill same-contract mispricings inside a few heartbeats.

Combinatorial arbitrage is a different animal, and that's exactly why 290 episodes survived where 7 didn't. Spotting it requires holding multiple related contracts in your head at once (a moneyline, a spread-equivalent, a multi-outcome combination) and checking that their implied probabilities are mutually consistent, not just individually plausible. That's a modest step up in sophistication, mostly about tooling and monitoring breadth rather than actual forecasting talent, but it's apparently enough of a step that most participants don't bother, especially in the chaos of a game's final minutes when lines are moving fastest and attention is thinnest.

#CAPPED BY DESIGN, NOT BY EFFICIENCY

Before anyone gets excited, the paper's own numbers undercut the opportunity almost as fast as they reveal it. Three out of four combinatorial windows were executable only at retail size: an average of 14.8 shares, which at typical NBA in-game prices is lunch money, not a strategy. The 101 basis point median return is real, but it's real for a trade size that can't be scaled without moving the book yourself and erasing the edge you came for. The authors' own conclusion is blunt: Polymarket NBA markets are close to efficient in aggregate, and the inefficiencies that remain are structurally confined to the retail tier, with diminishing returns for anyone trying to deploy real capital against them.

#WHAT THIS ACTUALLY MEASURES

Put the two studies side by side and you get the real lesson. The election markets had strong volume, weak accuracy (67% for Polymarket), and documented cross-platform incoherence, plus, per contemporaneous reporting, wash-trading allegations and a large anonymous "Trump whale" position that make apparent order-book depth impossible to fully trust. The NBA markets had far less cultural attention but killed same-contract mispricings in 3.6 seconds flat. That's the opposite pattern you'd expect if volume and accuracy were the same thing, and it's exactly what you'd expect if internal consistency is a separate, capital-and-attention problem rather than a forecasting one.

So here's my actual claim: don't grade a prediction market on whether it nailed the outcome, because that's one sample and it's contaminated by luck. Grade it on how long a self-contradiction survives inside its own order book, because closing that gap requires no insight at all, just money and eyeballs watching closely enough to notice. Three point six seconds on Polymarket's NBA book says serious capital is there. A same-outcome probability split against Kalshi during the highest-volume event in the platform's history says it wasn't, not fully, not everywhere. Volume tells you people are betting. Internal consistency tells you whether anyone competent is actually checking the math, and that's the number I'd want before I trusted a price.

Calling the Future vs. Agreeing With Yourself

67Polymarket78Kalshi93PredictIt

In the 2024 election, Polymarket carried the highest trading volume of the three major platforms yet posted the lowest directional accuracy at 67%, versus 78% for Kalshi and 93% for position-limited PredictIt, the reverse of what efficient-market theory predicts from volume.

#Further Reading