Our model
The Parlay estimate
Prediction markets are good forecasters with a known flaw: in lopsided races, their favorites win more often than their prices say. The Parlay estimate corrects for it, using 125 decided races from 2024 and 2025 on Kalshi and Polymarket, and leaves competitive races at the market's odds.
How it changes the odds
Races where the favorite is under 70% keep the market's odds; the adjustment phases in between 70% and 80% and applies fully above that. It never flips a favorite.
| Market says the favorite wins | Parlay estimate |
|---|---|
| 60% | 60% |
| 70% | 70% |
| 75% | 80% |
| 80% | 90% |
| 85% | 94% |
| 90% | 97% |
| 95% | 99% |
How it was tested
We collected every decided Senate, House, governor and presidential-state race we could find on both exchanges, with daily prices from up to 90 days before Election Day. Each race was scored by a version of the model that never saw it (five-way cross-validation, repeated), and compared with the raw market odds.
| On races the model never saw | Market odds | Parlay estimate |
|---|---|---|
| Log loss (lower is better) | 0.239 | 0.220 (8% better) |
| Brier score (lower is better) | 0.0756 | 0.0741 |
In decided races where markets priced the favorite between 65% and 80%, favorites won about as often as priced, which is why competitive races are left alone. We also tested price momentum, disagreement between the exchanges, thin markets, distance from Election Day and LunarCrush social data (each favorite's share of the conversation against its share of the odds, and sentiment, on 43 races with full social history). None improved accuracy on held-out races, so none are in the model yet. They will be retested on 2026's results, with several times more races, along with trade flow from our trade archive.
Where the estimate and the markets differ most
2026 races where the model's adjustment is largest. Every race page shows its Parlay estimate beside the market odds.
| Race | Market | Parlay estimate |
|---|---|---|
| Wisconsin Governor | D 81% | D 91% |
| OH-07 (Ohio) | D 81% | D 91% |
| TX-15 (Texas) | D 80% | D 90% |
| NY-01 (New York) | R 81% | R 91% |
| Texas Governor | R 81% | R 91% |
| PA-10 (Pennsylvania) | D 81% | D 91% |
| VA-02 (Virginia) | D 82% | D 92% |
| AR-02 (Arkansas) | R 82% | R 92% |
| IA-01 (Iowa) | D 82% | D 92% |
| NE-01 (Nebraska) | R 82% | R 92% |
Questions
- What is the Parlay estimate?
- Parlay the People's own probability for each race, built by adjusting the prediction-market odds with a model trained on how decided races actually turned out.
- How is it different from the market odds?
- In decided races, clear favorites (priced at 80% or more) won more often than their prices implied. The model corrects for that. Competitive races, with the favorite under 70%, keep the market's odds.
- Is the Parlay estimate betting advice?
- No. It is election research. Parlay the People does not make buy or sell recommendations, and what readers do with the research is their own decision.
- Does social media buzz predict winners?
- Not on its own, so far. On 2024's decided races, a candidate getting more of the online conversation than their share of the odds was not more likely to win. The Pulse shows it because it often explains why a market is moving, and we will keep testing it as data grows.
- How will we know if it works?
- Every estimate is saved hourly and scored after Election Day against what the markets said, with the Brier score and log loss. The results will be published here, win or lose.
