# Better prediction models with polling data?

I've been working on a project on measuring polls' accuracy in complex contexts (more than two candidates) where there are a small number of inaccurate polling data points.

I thought it would be the case to work on the probabilities instead of focusing on the real differences between the estimated and the actual vote.

For this, I thought designing a Kalman filter to filter out any noise in the data, and then predict daily data backwards between the day of the election until the appearance of the first poll. After this, I thought on track the probabilities of a candidate win or advance in the runoff stage, so I can plug in the simulated data by the Kalman filter into a Dirichlet-Multinomial model to estimate those probabilities.

My question is, how compelling is this design? I think it might be another superior approach for this purpose, so I'd like to hear those.

• "I think there many other superior approaches for this purpose" - like what? If you have specific alternatives in mind, please edit the question to identify them.
– AirThomas
Jun 23 '14 at 18:18
• What do you mean by "inaccurate"? Imprecise (have a large margin of error)? Biased (systematically under- or overestimate the proportion for intending to vote for a given party)? Both? Kalman filtering will probably be able to deal with the former but not with the latter. Apparently, a good model was used by Nate Silver, so you just need to reverse-engineer it :) Jun 27 '14 at 13:32
• @StasK By inaccurate I meant 5 to 7 percentage points difference in the estimation with standard MoE 2 to 3 points. For instance, in the eve polls before the election, I have 2 pollsters overestimating candidate A in 5%, while underestimating candidate C in 7% and candidate B in 4%. If I cannot trust on the estimates even for a day before the election, how could I trust on them all the way back for a year or so? Jun 27 '14 at 18:45