US Open 2026: Were Prediction Markets Right?

A Brier-score analysis of 254 singles matches finds that Kalshi’s prices beat a 50/50 benchmark, while favorites won more often than their odds implied.

The US Open gives us two ways to judge a prediction: what happened on the court, and what traders thought would happen before the first ball was struck.

At the 2026 tournament, Alexander Zverev and Elena Rybakina won the singles titles. But knowing who lifted the trophies does not tell us whether the markets did a good job forecasting the matches that got them there.

For that, I returned to the approach used in my March Madness analysis: comparing pre-match probabilities with results using the Brier score.

Across all 254 main-draw singles matches, Kalshi’s prices produced an overall score of 0.1642. A forecast assigning a 50% chance to each player in every match would score 0.25. Lower is better.

Did the Prediction Markets Get the US Open Right?

A favorite losing does not automatically make a probability forecast wrong. A player given an 80% chance should still lose about one match in five across a sufficiently large set of comparable forecasts.

Poker players know the distinction. Losing with the best hand does not mean the hand was behind when the money went in. Equally, one unexpected winner does not establish that the market misunderstood the matchup.

The useful question is how the probabilities performed across the tournament, including the routine wins that attracted less attention than the upsets.

What Is the Brier Score?

The binary Brier score measures the squared distance between a forecast probability and the result. A win is recorded as 1 and a loss as 0. Those squared errors are then averaged across matches.

Brier score = average of (forecast probability − outcome)²

For illustration, a player given a 90% chance who wins contributes 0.01: (0.90 − 1)². If that player loses, the contribution is 0.81: (0.90 − 0)². The score therefore penalizes confident misses much more heavily than uncertain ones.

This analysis uses Kalshi’s match-winner markets for 127 men’s and 127 women’s matches. I took the latest usable bid-ask midpoints at or before one hour ahead of each match’s recorded start time, then normalized the two players’ probabilities to total 100%. Each match receives equal weight.

Qualifying matches and four listings with non-binary settlements were excluded. The remaining records form complete main draws, with each round’s winners matching the entrants in the next. These are Kalshi results, rather than a combined assessment of every prediction-market platform.

How Did the Markets Perform?

The combined score represents 34.3% less squared probability error than the 50/50 benchmark. That is evidence that the prices contained useful information about who was more likely to win. It is not an accuracy percentage or a measure of trading returns.

The women’s draw produced the lower score. The second round was particularly favorable to the markets: 31 of 32 favorites won, producing a women’s second-round Brier score of 0.0675.

Breaking down both draws together shows that the scores did not simply worsen as the tournament progressed.

The semifinals scored well, with all four favorites winning. The two finals, taken together, finished slightly worse than the 50/50 benchmark. With only two matches, that is a description of those results, not a reliable verdict on how well markets price finals generally.

The timing check also matters. Using prices 15 minutes before the recorded starts produced a score of 0.1643; five minutes before produced 0.1644. The overall finding barely changed.

Which Results Hurt the Markets Most?

The largest individual error came from Dane Sweeny’s second-round win over Lorenzo Musetti. Musetti had been assigned a 93.5% chance in the one-hour snapshot. His defeat contributed 0.8742 to the score before averaging.

Other notable misses included Novak Djokovic’s opening-round loss to Mariano Navone and Amanda Anisimova’s third-round defeat by Anastasia Potapova.

The women’s final was a smaller surprise. Aryna Sabalenka was priced at 57.5%, giving eventual champion Rybakina a 42.5% chance. Zverev, meanwhile, was approximately a 57.9% favorite against Ben Shelton in the men’s final. Neither championship match was priced as a foregone conclusion.

Final Verdict – Were the Markets Right?

The prices performed meaningfully better than treating every matchup as a coin flip. Higher-priced favorites also won more often than lower-priced ones: favorites in the 90%–100% range won 24 of 26 matches, or 92.3%, against an average forecast of 92.9%.

But a good aggregate Brier score does not, by itself, establish perfect calibration. Across the 253 matches with a clear favorite, the prices implied roughly 183.5 favorite wins. There were 197. One match was priced at exactly 50/50 and is excluded from that favorite count.

That suggests the favorites did better than the market probabilities anticipated in this tournament. It does not establish a lasting tendency to underestimate favorites; a single event remains a limited sample. Nor does the lower women’s score prove that those markets were intrinsically more accurate.

Scores depend partly on how difficult the matchups are to forecast and which outcomes happen. A comparison with sportsbooks or another forecasting model would require probabilities for the same matches at comparable times.

The conclusion is narrower, but still useful: Kalshi’s US Open prices provided informative forecasts across both draws, even though several heavy favorites lost. The biggest upsets are part of that assessment. They are not the whole assessment.

Methodology

Original calculations from Kalshi event and milestone records and one-minute candlestick histories, retrieved Sept. 18, 2026. Main-draw singles only. The primary cutoff is 60 minutes before the start timestamp in Kalshi’s milestone record; those timestamps have not been independently verified against first-ball logs. For each player, the most recent usable quote at or before the cutoff was used, allowing up to 10 minutes of staleness. The maximum observed age was seven minutes and the maximum bid-ask spread was two cents.

For each player, midpoint = (Yes bid + Yes ask) ÷ 2. Player A’s probability = midpoint A ÷ (midpoint A + midpoint B). Score = (probability A − outcome A)², averaged once per match. The other player gives the same score. Results follow Kalshi’s binary settlements, including retirements resolved under its match rules. Fees and trading profitability are not assessed. Figures are rounded for display; calculations use unrounded values. Supporting data retain raw responses, source URLs, exclusions and reproducible calculations.