For a given question we pick forecasters who have beat the crowd on all related questions. Then we weight each by their sum of accuracy on the related questions.
Let the forecaster weight be the Y axis and forecast from 0 to 100 be the axis. Now divide predictions into 2 clusters on this plane. Sum the weight of each forecast in a cluster. The median forecaster value of the cluster with the highest sum of forecaster weights is the forecast!
I can think these up all day long. They’re probably all equally nonsensical. In this case I am looking for an application of SciPy kmeans2. This will also give me the opportunity to generate some authoritative-looking pictures of my nonsense, along these lines: