The "Monkey Pollster" vs fivethirtyeight.com

There's a lot that went into fivethirtyeight.com's state-level polling analysis. But I wondered if a very basic approach to poll aggregation would perform just as well - in much the same way that the simple Marcel "Monkey" baseball predictions perform just as well as the more advanced projection systems.

The "Monkey Pollster" simply averages the last five polls before the election; if there were fewer polls, it uses only them; if there were multiple polls on the same day and it's only possible to determine what the last five polls were, they all get used. Here's how the "Monkey pollster" did compared to fivethirtyeight.com:

R^2 of actual Obama margin of victory vs model:

538 - 0.9704; Monkey - 0.9686

[Uses line of best fit with non-zero constant...]

Mean-Squared Error:

538 - 35.7; Monkey - 28.3

Electoral Votes: (365 Actual)

538 - 353; Monkey - 338.

Both systems missed Indiana, and the Monkey prediction was 0.7% low for North Carolina, while fivethirtyeight was 0.7% high - in the right direction. Overall, the Monkey performed just as well as the more sophisticated system.

Let me be clear that this is in no way an indictment of the work that appeared on fivethirtyeight.com. I think it was brilliant to incorporate Monte Carlo analysis into the presidential election - but we can also see that a complex poll-weighting system may be no better at rejecting outlying data than a simple averaging, even of bad polls.