When referring to elections, “Polling Average” is a commonly-heard term. How does the World Election Forecast calculate polling averages? Before we begin, we must answer the question: “What is a Polling Average?” In short, it is an weighted average of all opinion polls regarding an election. These polls are weighted on the basis of recency, polling accuracy, and sample size.
Every time a new poll comes in, the last date of polling, sample size, link (for reference purposes only) and the results for each of the major candidates (in the example shown below, we show results for the 5 major French Presidential Election Candidates) are recorded. The results are input into a spreadsheet, which calculates the weighted value according to the following formula:
Weighted Vote = [8 * (Sample Size)^.7] * [0.7^(Days since Poll)] * Actual Vote
In the first section [8 * (Sample Size)^.7], the poll is weighted based on its sample size: larger polls have a stronger weighting than smaller ones, but the effect is not linear. By the formula, a poll with 1000 respondents will have a weighting of about 1007, a poll with 2000 will have one of about 1636, and a poll of 15000 will have about 6704. By weighting polls as such, we try to minimize potential errors in a single poll – if a poll with 15000 respondents is flawed slightly, the polling averages will not be as heavily affected. However, the weighting allows larger polls to retain their ability to be weighted more than smaller polls, because in theory they are more accurate.
In the second section [0.7^(Days since Poll)], the poll is weighted based on the number of days from the present. This is more obvious – an earlier poll will be less obviously accurate. However, we cannot simply disregard all polls not on the day of measurement – in many cases day-to-day fluctuations are merely statistical “noise” and not actually significant. By this weighting measure, polls a day old are weighted normally, polls three days old are weighted at about half, polls a week old are about a tenth, and polls two weeks old are at about 1%. Note that base of the exponential function does not have to be 0.7 – it varies depending on the availability and frequency of polling.
There are additional auxiliary measures that are measured after the value has been computed. First is the day-to-day fluctuation error, which is the measure of the difference between the cumulative results of each day. This is used to calculate the volatility and trend of a candidate. The weighted volatility is the average day-to-day error of each candidate multiplied by the 2.5th logarithm of the days until the election. This measure is in order to determine the degree of fluctuation each candidate is expected to undergo, and is a major part of our forecasting model.
After the finalization of a polling average, the measures are input into our model to compute the final probabilities.
Polling Average Example (in excel format)