Thursday, January 19, 2017

Risk mining

The above is the FiveThirtyEight Premiership win probabilities. The probabilities are in the right column. The current positions and point totals are in the left column. One could wonder what is the value added of a model that's converting the observed form -- as reflected in the tables -- into a probabilities that reproduce the ordering of the table.

That conceptual challenge of probabilistic forecasting, but more importantly, the deceptive ease of applying the methodology to political forecasts, is something that 538 impresario Nate Silver is not much interested in discussing. He is however working towards the conclusion that the need to rethink election analysis is deepest at the political desk of the New York Times. That's despite the fact that the justification of his poll-driven approach is heavily reliant on the notion that certain news stories get reflected in poll movements -- but he has no theory of which news stories make that breakthrough. In football, the events are clear: matches and scores. In political campaigns, the field, rules, and spectators are ill-defined, and there's only one actual electoral event. With the vast amounts of potentially relevant information and a single event, a slippage from data to punditry is inevitable.

No comments: