Last Sunday, voters in Extremadura made their choice, bringing pre-election polls back into the spotlight. This time, polls predicted a confident win for the Partido Popular, but without an absolute majority. However, the details of the vote distribution and support for other parties took many by surprise. The average error for the parties was about 2.4 percentage points—slightly higher than usual. Compared to previous years, when forecasts were more accurate, this year’s results left experts with questions.
In 2024, polls showed much greater accuracy, especially during elections in Catalonia, the Basque Country, and Galicia. Back then, analysts noted the average margin of error did not exceed two points. This time, even pooled data failed to significantly reduce the error. Polls correctly identified the winner and noticed the decline in PSOE’s support, but underestimated the extent of that drop. Vox, on the other hand, was undercounted by nearly three points. In the end, Extremadura’s voters leaned further to the right than expected: together, Partido Popular and Vox received 60% of the vote—something no poll predicted.
Margins of error and probabilities
The probabilistic forecasting model proved its effectiveness: it predicted in advance the most likely outcome—a victory for Partido Popular without a majority, estimating the probability at 88%. Wide probability intervals explained why the final result was entirely possible, even with imperfect polling. All parties, including PSOE and Vox, won the number of seats that fell within the previously defined ranges with an 80% probability.
Experts emphasize that election predictions always involve an element of chance. It is incorrect to judge the work of polling organizations based on a single case—what matters is their accuracy over the long term. A comparative analysis of eight leading agencies since 2018 shows that GAD3 and 40dB have consistently delivered the best results. Average data also indicate high accuracy, but this is mainly due to the quality of the original polls.
Problems with government polls
The latest estimate by the state institute for the elections in Extremadura showed an error of 2.9 points per party—worse than most private agencies. The forecasts for Vox and Unidas por Extremadura were accurate, but Partido Popular was underestimated by four points, and PSOE was overestimated by six. Nevertheless, in terms of the number of seats, this forecast was among the three best, only behind Sigma Dos. On average, the deviation per party was 1.3 seats, and for Sigma Dos it was 1.1. The average data missed by about two seats for each political force.
Since 2018, the state institute has remained the least accurate of the eight agencies analyzed. In 11 out of 13 elections, its error exceeded the average level, and in five cases it had the worst result. This is confirmed by a comparison of party errors by percentage of votes.
Systematic bias
This time, the main problem of state polls resurfaced—a persistent bias in favor of left-wing parties. The institute projected that PSOE and Unidas por Extremadura would receive 42% of the vote, but in reality they garnered only 36%. Experts are no longer surprised by this six-point gap: a similar trend has been observed since 2018. Over this period, in 42 out of 43 elections, the institute overestimated support for leftist parties.
This scenario repeated itself in the 2023 parliamentary election, in all four campaigns of 2024, and now—again in Extremadura. Despite recurring errors, there have been no methodological changes, and the bias persists year after year.
