Forecasting an uncertain event · Contested (since 2026)

On 16 July 2026 the Forecasting Research Institute reported that several systems on ForecastBench could no longer be told apart statistically from its panel of superforecasters, and that one of them ranked above the median superforecaster on market questions for the first time.

The benchmark asks questions whose answers do not exist yet, which is what keeps a model from having read the outcome somewhere. When it was published in 2024 the expert panel beat the best model with a p value below 0.001. What happened in July 2026 is a set of overlapping intervals rather than a defeat, and the institute says as much: the human forecasts it compares against were collected in 2024, so the gap is being closed against a panel that has not played since, on an extrapolation that ages. Contested is the honest word, and it will stay that way until the panel forecasts again. What a forecast is for sits outside the board anyway. The question nobody thought to write down, the decision taken on the number, and the person who answers for it when the number was wrong are three things no Brier score has ever measured.