Nick Firoozye writes:
In Greece there are 10 polling companies with very different methodologies. There are some 15 parties but 3 of any major importance (ND, Pasok and Syriza). The relevant polls that were done involved one before the election, and two after. We also have the outcome of the most recent election (so we know something about error). Each polling company has biases. Some are far more accurate than others. Most seem to use some stratified sampling method, but some use an ARIMA to adjust for respondents’ recent polls. Dimitris can go into detail on the differences of methodology but some of this may be unimportant for purposes of coming up with a better estimate of voting outcomes.
Dimitris has also pointed out a poll combination method he read at 538, but this seems overly complex for the data available.
Some may have better methodology but it would appear more reasonable to incorporate more data and do some weighted average. We were wondering if sort of problem this is of interest to you and what you would recommend for a method for how to weight the polls?
The polls themselves do not produce ‘forecasts’ since they are meant to be contemporaneous (and I guess trend matters then, although with the limited timeseries data, we may have to be happy with contemporaneous estimates). As you can tell it is awfully pertinent now.
My reply: do some time series plots with a dot for each poll, I don’t know offhand how much better you can do than to draw a line by eye. But others might have more informed thoughts on this?