How to read odds and forecasts

A probability is an estimate

If a model gives one side a 62% chance, it is saying that if the same conditions played out 100 times, that side would win about 62 of them. The other 38 are not a small number.

A probability is not a vote share either: a 62% chance doesn’t mean winning 62% of the vote.

Why do models differ?

• They use different data: polls, past results, economic indicators, expert ratings.
• They weigh the same data differently: which polls to trust more, or how much an incumbent’s advantage counts.
• They update on different days: a new poll may be in one model and not yet in another.

So we don’t hide the difference by averaging: each model sits on its own row, with its own date.

Why dates matter: the 14-day rule

Probabilities go stale. Every row shows the model’s publication date. A result older than 14 days appears faded, with an “older than 14 days” note, so you don’t mix up a current estimate with an old one.

What prediction-market prices are

• Polymarket is a prediction market: traders buy and sell contracts that pay out if an outcome happens. A price of 62 cents is often read as a 62% chance.
• Prices reflect the people trading and the money behind them. In thin markets a few trades can move the price a lot; check the volume shown in the widget.
• Prices can move within minutes, especially on election night. The widget shows Polymarket’s current price; we don’t store or average it.
• We show prices through Polymarket’s own embed, with buy buttons turned off and no affiliate links. Prediction markets aren’t legal or available everywhere, and showing prices isn’t advice to trade.

Ratings are not probabilities

Cook Political Report, Sabato’s Crystal Ball and Inside Elections sort races into categories: Toss-up, Tilt, Lean, Likely and Solid. They don’t give percentages; they are expert assessments. We show each rating with its date, and a history when it changes.

We don’t have our own model

US Midterm 2026 doesn’t produce its own forecast model or an average of models. We pass on only public, dated, linkable model results, with their source.

On election night, results speak, not probabilities: “Won” appears only with an AP or Decision Desk HQ call, with source and time.