Pollster ge’s forecasts are based on the aggregation of publicly available polling data.

What kind of data are we using?

We use those public opinion polls in our forecast where respondents were asked about their voting intentions, party identification, or political sympathies. Since the October 2020 parliamentary elections, we no longer use opinion polls conducted for or by particular political parties. We also include results of proportional voting in elections, namely party-list polls for parliamentary and municipal elections.

How is data aggregation done?

We aggregate our data using time-series statistical models, specifically generalized additive models (GAMs). Apart from the key explanatory variable, time, we also control for a polling initiative, interview mode (whether interviews were administered over the phone or face-to-face). Apart from polling information, our model also uses economic and epidemiological information, such as lari exchange rate, monthly inflation rate, the number of new COVID cases and deaths.

How do we control for the quality of a poll?

As a result of the 2020 parliamentary elections, we came to the realization that various polls estimate political sympathies and voting intentions with different accuracy. Thus, for our forecasts, each of the polls will be assigned a weight. The model will consider each of the results proportional to their quality weight. For polls conducted after the 2020 parliamentary elections, weights account for how accurate a survey was relative to the parliamentary polls. Studies conducted before the 2020 parliamentary elections are weighted based on their performance in the 2018 presidential elections.

In the table you can see the weights for each of the polling initiatives. Election results have the highest weights, followed by studies done by Edison Research and Survation.

Replication code and raw data for our forecasts are available from here