Polymarket had Bolsonaro at 77%. Polls had it 50/50. Lula won.

I build a trading bot for Polymarket, so Brazil's runoff on Sunday was an interesting one to watch. The week before the vote, the market priced Flávio Bolsonaro at 77% while Datafolha and MDA had the race basically tied. On election night, after 8:30 p.m. ET, Kalshi still had Lula at 21%. He won 51.8 to 48.2.

The next morning B3 hit its circuit breaker at the open for the first time since 2020. The Ibovespa closed down 13.9%, the dollar jumped 8% against the real, and EWZ lost 22% in New York. Stocks had been priced for a Bolsonaro win too, and the market moved all at once when that turned out wrong.

My takeaway as a builder: my bot estimates probabilities with Bayesian updates, and the easy shortcut is to treat the market price as the prior. In this case the price and the polls disagreed for three weeks, and the price was the one that was wrong. A 23% outcome happens about one time in four, so this alone doesn't prove the market was broken. It does make me less willing to let a crowded political market override the polls.

If you've built anything on prediction market data: do you trust the price over the polls, or treat a gap between them as a signal?