David M Rothschild on Posted on

The topline probability of the Democratic candidate to the win the presidential election has several different types of data, but is mainly driven by prediction market data.

National Data: we follow market contracts on several exchanges for the Democratic and Republican candidates to win the election. We have been following those contracts for years. From the start of the contracts we: de-bias, normalize, and average. The de-bias eliminates favorite-longshot, the normalization assures that Democratic + Republican = 100%, and then we average. For the National Forecast Only we do not use any fundamental or polling data.

The de-bias procedure is outline and validated here in depth, but can roughly be translated by: normal{1.64*inv.normal(price)}.

We do not use bookie data in the average, because it is only one side of the market. It is what someone is willing to sell the contract for, but not what people are willing to buy the contract for. We do not use Iowa Electronic Market, because it is technically for popular vote, not winning the election (and it lacks liquidity).

State-by-State Data: we follow market contracts for each state along with polling and fundamental data. While we launch state-by-state forecasts in mid-February, by mid-May we have state-by-state predictions that we really believe provide information above and beyond the national forecasts. Here is more detail than you want on the state-by-state forecasts, but here is what you want to know. (1) We make a separate forecasts for each of the three data types. The prediction market forecast, like the national forecast, takes the contracts: de-bias, normalize, and average. The polling forecast takes the aggregated polling average and projects it forward to Election Day by taking into account historical trends for incumbent party candidates. Then, using the standard deviation of the expected voteshare error, creates a probability of victory. The fundamental model is explained in detail here, but includes: presidential approval, economic indicators, past election results, incumbency, and home states. (2) We average the forecasts, starting with a very fundamental heavy forecast in February to a very market heavy forecast by Election Day.

Of course, we need to aggregate the states to create a forecast of who will win the Electoral College. We do this by assuming that events that occur between now and Election Day are very highly correlated, but that Election Day error is slightly less correlated. We create a matrix of correlation for both of these with past: prediction market data and polling, along with overlapping demographic data. We run 100,000 simulations of the election every day to create the state-by-state probability of victory. Which is updated daily here.

PredictWiseForecast: average of national and state-by-state forecast. You can see below that they track each other very closely, but the state-by-state forecasts have been more optimistic for the Democratic (leading) candidate.