Somewhere inside a 273-row spreadsheet, four trades have gone missing.
I caught it on the second read, the way you catch a checksum that almost balances. A new scorecard from the on-chain data firm Bitquery reports that backing the favorite in Polymarket's 2024 primary elections would have lost about 4% per dollar staked. Buried inside that number is a stranger one: the favorites won. 238 of 273 races resolved the way the market expected โ an 87.2% hit rate. Then the audit splits the sample. 182 markets where the favorite traded above 90 cents, of which 177 resolved correctly, a 97.3% win rate. And 91 markets in the 50-to-90-cent band, winning at roughly 71%. Add the buckets back together and you get 242 winners. Not 238. And since 182 plus 91 exhausts all 273 markets, there is no third bucket sitting in the appendix holding the missing four.
Four wins, gone. It is a rounding error โ or it is the seam where the real analysis begins. The headline is a 4% loss. The actual finding is that the loss is not distributed the way anyone assumes.
What the audit actually measured
Polymarket is the largest prediction market running on-chain. It settles trades on Polygon, operates a centralized orderbook and front end, and โ this matters โ carries no native token. There is no emissions schedule, no farming curve, no governance coin whose price you can watch to infer sentiment. What it sells is information: binary contracts on real-world events that pay $1 if the outcome hits and $0 if it doesn't. Buy a "Yes" share at 90 cents and you are buying an implied 90% probability.
The Bitquery scorecard asked a narrow, uncomfortable question. If you had mechanically bought the favorite in every 2024 primary market โ no judgment, no edge, just the higher-priced side โ what would you have earned? The answer, per the audit, is negative four cents on the dollar.
But the methodology behind that number is where I slow down, because I have run backtests like this and I know how easily a clean result hides a dirty one. The audit prices each contract using a 24-hour average ending at 12:00 UTC on voting day โ a short-window TWAP that a single large trade can drag around. Where no trade occurred in that window, it falls back to the last trade within 30 days, which means some "prices" in this dataset could be a month stale. Runoffs are handled at runoff-day prices, which is reasonable. What is not reasonable, and what the authors openly concede, is that the -4% figure does not fully adjust for bid-ask spread, slippage, or fees. The reference price is not a fillable price. Which means the true cost of this strategy is worse than the number being reported โ the -4% is a floor, not a ceiling.
And the data itself is narrower than it looks. The study pulls from Polygon only. It excludes Polymarket's US-facing app and excludes Kalshi entirely. Nine markets were dropped from the sample โ early settlements, missing prices, one market with no settlement record at all. So whatever this audit proves, it proves something about one settlement chain, one venue, and a slightly truncated sample โ not about prediction markets as an industry.
The number hiding in the buckets
Now reconstruct the economics, because this is the part nobody has done yet.
If 238 favorites won out of 273 and the strategy lost 4% per dollar, then the average entry price across the whole book works out to roughly 90.8 cents. That is the blended cost. But favorites do not trade at one price โ they trade across a range, and the range is where the money lives.
Take the above-90-cent bucket. 177 wins out of 182 markets is a 97.3% hit rate. To lose 4% overall while this bucket wins 97% of the time, its average purchase price has to sit near 97.7 cents. Run that expectation: 0.973 multiplied by a $1 payout, minus 0.977 multiplied by a $1 cost, lands at roughly negative half a cent per share โ essentially break-even, before fees.
Now the middle band. 91 markets at a 71% win rate, average entry of 77 cents. Expected value: 0.71 times $1, minus 0.77 times $1 โ about negative 6 cents per share, or roughly -7.3% ROI.
Stack them and the blended -4% is not a uniform tax on "backing favorites." It is almost entirely manufactured by the 50-to-90-cent band. The ultra-favorites above 90 cents were, on this data, a near-arbitrage that barely bled. *The naive lesson โ "buying favorites loses money" โ is wrong. The correct lesson is that buying marginal favorites loses money, while the crowd prices near-certainties almost exactly right.*
The contrarian reading
Which flips the story on its head.
The popular takeaway from this audit will be that prediction markets are inefficient and that favorites are overpriced. The data says something more uncomfortable: at the top of the probability curve, Polymarket is remarkably efficient โ so efficient that 97% accuracy yields no edge at all. The inefficiency, the money actually left on the table, sits in the ambiguous middle, where 71% of favorites win but trade at 77 cents, a gap the market simply refuses to close.
And that gap is a trap of its own. That -7.3% is a pre-fee number. Polymarket's political markets currently charge somewhere between zero and one percent. Add a point of friction to a bucket that clears at half a cent of margin and it goes from break-even to bleeding. The near-arbitrage in the above-90-cent tier is not a strategy; it is a rounding error waiting to be eaten by costs.
There is a second blind spot, deeper than the first. This audit measures pricing. It says nothing about settlement. Event contracts ultimately resolve through an oracle or a centralized adjudication process โ and the study never touches that layer. One market in the dataset lacked a settlement record entirely. That single missing row is not trivia. It is a hint that the on-chain indexing layer does not fully capture how these contracts actually resolve, and that anyone backtesting prediction markets against a commercial data feed is building on an incomplete map. You can be right about the price and still be wrong about the payout โ and that risk never appears anywhere in this dataset.
Finding the signal in the static of the new wave
Here is what I take from it, in a market where survival beats returns.
The prediction-market narrative โ wisdom of crowds, priced in real time โ is real, but it is not a free lunch, and it is not evenly distributed across the probability curve. The wisdom is sharpest exactly where the payoff is thinnest. The opportunity, if it exists, hides in the murky middle, and it hides there precisely because the cost of harvesting it is high enough to keep most capital away.
The question worth asking is not whether Polymarket is efficient. It is whether an edge that survives fees, slippage, and a stale-price fallback can ever be large enough to matter โ or whether the real product here was never the trade, but the signal the trade leaves behind in the static.
Because the four missing wins were never really missing. They were a reminder that when accuracy and profit diverge, the spread between them is where the story lives โ and finding that signal in the static of a new wave is the whole job.