Visualizing clusters of current forecasts. Visualizing clusters weighted by forecaster category accuracy.
Decaying importance of old unupdated forecasts. Actually I don’t think I need to decay because filtering against the median plus a standard deviation has the same effect.
Measuring correlation of accuracy scores on GJOpen assigned categories. Producing synthetic categories (principal components?) for closed questions. OK now that seems easy, just make a time series of forecasts for each question and then do PCA on that time series to produce some basis vectors. The basis vectors are then the categories. (Now I’m talking shit, right? It sounds good but I’m not sure I believe myself listening to it. I have to actually try it instead of talking about it, amirite?)
Game-based models and Swedish logic. Any other old numerical model including ones proposed by other forecasters. Simulation-based models. Models driven by RSS feeds.
Something involving one-shot Bayesian Program Learning. (이봐! That’s a joke!)