Did Prediction Markets Read Brazil’s Election Count Too Slowly?

A count is not a random sample

Early election returns can mislead when places with different political preferences report at different speeds. A candidate trailing in the published total might still be likely to win once the remaining votes arrive.

In Reading the Count: Prediction Markets and Brazil’s 2026 Presidential Election, Raphael Coelho of the São Paulo School of Economics, FGV, compares election-count projections with trading on Kalshi and Polymarket. His central claim is that the markets were slow to absorb evidence favouring Lula during the runoff, although that conclusion depends on the projection used for comparison.

The preprint reports that Lula defeated Flávio Bolsonaro with 51.79% of the two-candidate vote on 25 October 2026. All election-night times below are Brasília time.

Reconstructing the information available

Coelho uses electoral court records covering 499,205 polling sections. These include vote totals and timestamps for printing and receiving tally sheets. He reports that ordering sections by receipt reproduces the national count, allowing a minute-by-minute reconstruction of which sections were public.

The study compares this reconstruction with trades and prices from both markets. Winner contracts pay according to the election outcome; their prices serve as market-implied probabilities, though fees and trading risks can affect that interpretation.

Two statistical benchmarks play different roles. The website Arvor published a live projection every five minutes beginning at 19:15. It compared municipal runoff results with first-round results at corresponding stages of counting.

Coelho’s own projection instead compares reporting sections with their first-round votes, then estimates results for sections still outstanding. Crucially, it uses detailed records released after election night. It measures information contained in the count, not a forecast traders could necessarily have calculated live.

The reported gap

At 19:15, with 1.6% of sections counted, the live projection assigned Lula a winning probability of 99.8%. Both markets reached a sustained 90% probability at 20:24, when 46% of sections were counted.

Coelho translates that difference into an uncertainty comparison. If markets shared the live projection’s expected final margin, their prices would imply a standard deviation roughly four times as large. Standard deviation here measures the spread of possible final margins around the estimate.

That interpretation is conditional. A price alone cannot reveal whether traders expected a different margin or attached more uncertainty to the same margin. The paper explicitly acknowledges that its ratio combines those possibilities.

The benchmark also matters. Tripling the live projection’s uncertainty delayed its sustained 90% threshold to 20:23, almost matching the markets. A simpler model, which extrapolated each state’s remaining votes using its current vote share, reached 90% at 20:36.

Profits depend on the forecast

The paper values actual purchases against each benchmark’s estimated winning probability. Using the live projection, Lula purchases from 19:15 had an estimated expected profit of 18 cents per Kalshi contract after fees. Tripling its uncertainty reduced that estimate to about four cents.

Using the simpler state-level extrapolation, expected profit was approximately zero. These are model-based valuations, not evidence of a risk-free opportunity. They also differ from realized profits calculated after Lula’s reported victory.

A suspension and a counterexample

Publication of the count stopped from 17:07 until shortly after 19:00 following an extension of voting at sections still open. During the suspension, markets priced Lula between 25% and 44%.

Retrospective section-level projections favoured Lula, assigning probabilities between 57% and 92% across calibrations. But those calculations were not available live. Arvor’s live method did not project below 1% counted, and printed tally sheets were not centrally collected. The suspension therefore offers weaker evidence of an actionable missed opportunity.

First-round night provides another qualification. The eventual leader, Flávio, led the headline count from its first release. Markets reached 90% for him about an hour before Coelho’s section-level projection. That projection’s historical calibration was also poorly matched to the election.

What remains uncertain

Forecast uncertainty is a central limitation. Calibration on the 2022 runoff produced bands that were too narrow in 2026. The alternative 2018 calibration required simulated reporting orders because receipt timestamps were unavailable, and checks suggested those simulations could understate errors.

Coelho discusses language barriers, platform blocking in Brazil, coarse forecasting methods and trading constraints as possible explanations, but cannot test them. The study supports a narrower conclusion than general market failure: on the reported runoff night, detailed projections identified the winner earlier than market prices, while simpler models and another election night yielded a different comparison.