What a 70 percent chance really means, checked against 479 races

When a forecast gives a candidate a 70 percent chance, it is not saying the candidate will win. It is saying that if you took every race where we gave a favorite about 70 percent, roughly seven in ten of those favorites should win and three in ten should lose. A forecast that never sees its 70 percent favorites lose is not brilliant. It is underconfident, and its numbers should have been higher. A forecast whose 70 percent favorites lose half the time is overconfident.

That is a testable promise, and the 2026 midterms gave us 479 Senate, House and governor races to test it on. We sorted every race by the chance our final forecast gave the favorite and counted how often each group of favorites actually won.

The coin flips behaved like coin flips

The first row is the one people tend to misread. We had 33 races where neither side was better than a 60 percent favorite. Those are close to coin flips, and the favorite won 58 percent of them, almost exactly the 55 percent we expected. The 14 upsets in that group are not forecast failures. A forecast that called every close race correctly would be lucky, not good.

The middle was too cautious

The rows from 60 to 90 percent tell a more useful story. Across those 111 races we expected favorites to win about 77 percent of the time, and they won 88 percent. That gap means our probabilities in the lean and likely range were too low: races we called leaning one way were more settled than we said. Part of that is the year. A national wave pushes most races in the same direction, so favorites on the winning side of a wave come through more often than an average year would suggest. Part of it is the model, and we take it as a sign that our uncertainty was set a little wide.

The same lesson shows up in the ranges. Every race on the site carries an 80 percent range for the margin. In a well calibrated forecast about 80 percent of results land inside that range. In 2026, 87 percent did.

The safe races were safe

At the top of the table are 300 races where we gave the favorite at least 95 percent. Those favorites lost once. This matters more than it sounds, because safe races are where a model with a hidden flaw tends to embarrass itself. A forecast that gets the close races right by luck but loses a safe seat has a structural problem.

One number to sum it up

Forecasters often report a Brier score, the average squared gap between the probability given and what happened, where zero is perfect and 0.25 is what you get by calling every race a coin flip. Our 2026 score across all 479 races was 0.048. Across the whole table there were 28 upsets. If our probabilities had been perfectly calibrated we would have expected about 47. Fewer upsets than expected is the signature of the cautious middle described above.

How to read our numbers

• Under 60 percent is a toss up. Treat it as genuinely open, however the map is shaded.
• 60 to 90 percent is a real edge. In 2026 our edges held more often than the number said, but a single race is still a single draw.
• Above 95 percent means it would take an error well beyond the typical misses we measured in the 2018, 2022 and 2026 races to flip it.

The same reading applies to the 2028 forecast, and we will grade it the same way, in public, once the votes are counted.