Prior to the 2021 Municipal Elections of Georgia, Pollster.ge will again offer forecasts based on the publicly available polling data. Thus far, we have tried to predict the outcomes of the 2018 Presidential and 2020 Parliamentary elections. Some of our expectations proved to be accurate, with significant discrepancies. Before we present our forecast for the 2021 elections, we would like to briefly summarize the work we have done.

We have applied two different approaches to aggregate polling results, the weighted/running average method and “state space” statistical model. The first method implies simply averaging openly available survey results. Newer polls get larger weights, while older ones are getting smaller weights. This helps taking into account the fact that newer polls better reflect people’s attitudes close to elections.

The “state space” statistical modeling applies far more complicated procedures. In this case, we control for both the time factor and the fact that, presumably, a separate survey reflects only approximate public opinion. The state space model is based on Bayesian statistics. Our forecast is based on 2000 election simulations. Next, we take the median of the simulation percentages as a point estimate (or the most expected result). Simulated results also are used to calculate confidence margins. You can read more about pollster.ge methodology here.

Initially, we used polling averages to forecast outcomes of the 2018 presidential elections. To predict the results of the first round, we compiled the results of fifteen pre-election polls. As of October 25, 2018, the weighted average and “state space” models gave slightly different forecasts. According to the weighted average model, the candidate supported by the government, Salome Zourabichvili, should have earned 35%, while the state-space model gave her about 39%. In the first case, the forecast was off by 4% from the result of the first round, while the Bayesian forecast coincided exactly with the election. However, a significant difference was observed in the case of the United National Movement candidate. Both forecasts missed the election results by more than 10%.

Nonetheless, both methods indicated on the runoff. In the second round, predicting the election results were complicated by the fact that we had almost no new polling data. Before the second round, only a few polls were conducted, based on which it was clear that we would be dealing with the so-called horse-race elections. The weighted average evaluated the chances of both candidates in the second round equally, while the “state space” model gave a slight advantage to Salome Zourabichvili. As a result, the latter model was more accurate.

For the 2020 parliamentary elections, we had the results of more polls. However, standing where we are now, only diverse polling were not enough for accurate forecasting. We started publishing our predictions for the 2020 elections on July 13th and then updated our data 4 more times. We made the last two updates 2 weeks before the elections (October 14) and three days before the elections (October 28). It turned out that forecasts for October 14 were more in line with the final results than the predictions made before the elections. For example, in the case of the weighted average, the projected outcome of the ruling party differed by only one percentage point from the final results. As for the UNM, later forecasts were more accurate.

At the end, we bring up the comparison of the assumptions made by us three days before the elections with the final results published by the CEC, as well as with the four exit polls conducted on election day. Pollster.ge’s assumptions, however, differed significantly from the official election results, but on average, the error in this case is still less than that of some of the large-scale exit polls conducted on election day.

Based on the experience of the 2018 and 2020 elections, Pollster.ge was somewhat successful in predicting election results. Past experience has shown that outlier or biased polls might significantly skew forecasts. Here we also face the challenge of methodological transparency, where many opinion polls in Georgia often fail. That said, our future forecasts will take into account the following criteria:

• when modeling point estimates, we will minimize large-scale shifts in opinion that we do not account in simple average models.
• assess the accuracy of the polling initiatives compared to previous elections, by giving more weight to more accurate polls and vice versa.