Trusted Data
Operations For
The AI Era

Prediction markets are the newest tool in election forecasting. More than $21 million was traded on Polymarket's Texas Senate market alone. A contract pays $1 if its candidate wins and nothing otherwise, so a 74-cent price means about a 74% chance.

In theory, traders with money on the line pay closer attention than anyone. The 2026 midterms let us test that by checking daily Polymarket prices and poll averages against the results. To dig into the data yourself, explore the Senate and governors' races in the Midterm Forecast Ledger.

We should say up front that we aren't political scientists or statisticians, just interested observers who like data.

What we found

Polls missed, and mostly in one direction. In 30 Senate and governor races with a prediction market and at least five recent polls, the final poll average underestimated the Democratic candidate in 25. On average, Democratic candidates ran about 3.5 points ahead of their polls.

That gap shrank as Election Day approached but never closed. It was almost 6 points 90 days out, 5 a month out, about 4 a week out and 3.5 the day before. Prediction markets drifted at about the same pace. Across all 67 resolved Senate and governor markets on Polymarket, prices implied about 35 Democratic-aligned wins 90 days out and 39 the day before. The actual number was 42.

At Belvedere Labs we spend a lot of time trying out new AI capabilities, wherever they show up. Weather caught our eye because it makes an unusually clean test. The forecasts are public, the answers arrive a few days later from official thermometers and rain gauges, and two of the world's biggest forecasting agencies now run an AI model alongside their traditional one.

So we gathered the forecasts four models made for eight US cities from July to October 2026 and compared each one with what actually happened. This article is a short summary of what we noticed. The full data set lives in our interactive forecast explorer, where you can dig into any city, day or model.

Two Ways to Predict the Weather

Traditional forecasts come from physics. The model splits the atmosphere into millions of 3D boxes and applies the equations for air movement, heat and moisture, stepping forward a few minutes at a time until it reaches next week. That takes a supercomputer. We looked at the two best-known examples: IFS from the European centre ECMWF, and GFS from the US agency NOAA.

AI models take a different route. They are neural networks trained on decades of past weather, and they learn how the atmosphere tends to change over six hours. Repeating that step gets them a week or more ahead. Training one takes a lot of computing, but running it is far cheaper than running a physics model. ECMWF estimates its AI model uses about 1,000 times less energy per forecast, and Google DeepMind reported that its GraphCast model produces a 10-day forecast in under a minute on a single machine. The AI models still start from today's weather as measured and prepared by the traditional systems, so the two approaches work together. ECMWF made its AI model, AIFS, operational in early 2025, and NOAA followed later that year with AIGFS, which builds on GraphCast.

When I became a CIA operations officer, I wasn’t thinking about data pipelines.

I was thinking about people, information, and the mission. About understanding what was happening and getting the right information to the people who needed it.

But somewhere behind every useful piece of intelligence was a less glamorous reality: someone had to get the information collected, processed, connected, and delivered.

Nobody puts that part in the movie.

They probably should. Although “We finally resolved the schema mismatch” would be a difficult line to build a trailer around.

The People Behind the Information

In the field, I relied on talented people to help get me good information. I was fortunate to have them.

They needed reliable ways to ingest and prepare data so analysts could connect the dots and mission teams could act. That meant more than getting information quickly. It meant preserving context, understanding where it came from, and having confidence in what it actually told us.

A useful connection buried in an inaccessible system was still buried. Information that arrived after a decision had been made couldn’t inform that decision.

The people behind that work made an enormous difference. Often, their effort was invisible precisely because they made things work.