Parties sliding late in polls fell further 24 times in 29

When a party's support drops in the final weeks of a campaign, the last polls usually do not catch the whole fall. Across 65 national elections in nine countries since 1950, 29 parties or candidates finished the campaign at least two points below their average of the month before. On polling day, 24 of those 29 did worse again than their final polls showed.

Across all 263 party results in our data, the final polls were close to a coin toss: 140 came in below them and 123 above, with an average miss under a tenth of a point. Parties that rose by two points or more in the last month split almost evenly: 15 of 30 finished below their final polls and 15 above. The fallers are the group that stands out, finishing on average 0.9 points below their last polls.

This matters most in France, where round one of the presidential election is on 18 April 2027 and the race for a run-off place is tight. France is also the one country in our data with several such cases where the pattern has held without exception.

What we counted

We took every election outside the United States for which we hold both a poll history and a set of final polls: general elections in Britain, Canada, Australia, Spain, the Netherlands, Germany, Poland and Italy, and the first round of the French presidential election. For each party we compared two figures from our comeback and accuracy data. The first is the month-out average, which centres on 30 days before polling day and takes in each pollster's polls from two months out to the end of the campaign, with every pollster weighted equally except in Britain, where each poll counts once. The second is the average of each pollster's final poll. A party counts as sliding when its final polls sit two points or more below its month-out average. We then measured how its real vote compared with those final polls. Parties polling under 5 per cent at the end, and the "others" column, are left out.

The British cases include some of the best-known polling misses. In 2024 Labour's month-out average was 41.9 per cent; the final polls had it at 38.9 and it won 34.7 per cent in Great Britain. In 2001 the sequence was 48.0, then 45.2, then 42.0. In 1997 it was 49.9, then 46.6, then 44.3. The Dutch Socialist Party in 2012 shows the pattern at its starkest: an average of 19.5 a month out, 13.8 in the final polls and 9.7 on the day. In Canada in 2015 the New Democrats went from 27.5 to 21.3 in the final polls and won 19.7.

France's own record

Five French first-round candidates in our data slid by two points or more in the final month, and all five finished below their final polls. Lionel Jospin in 2002 averaged 20.3 a month out, 17.8 in the final polls and won 16.2, which left him out of the run-off. In 2017 Benoît Hamon went from 10.9 to 7.6 and won 6.4, while Marine Le Pen went from 24.6 to 22.2 and won 21.3. In 2022 Éric Zemmour went from 11.9 to 9.7 and won 7.1, and Valérie Pécresse went from 10.1 to 7.8 and won 4.8.

France differs from the wider sample in one respect. Its late risers also kept rising. Jean-Marie Le Pen gained two points in the final month of 2002 and beat his final polls by 3.9. Jean-Luc Mélenchon gained 3.1 points in the last month of 2022 and beat his final polls by 4.8, finishing about one point short of the run-off. All four French candidates who rose by two points or more did at least as well as their final polls. In the rest of our data, risers showed no such tendency.

Our projection for 2027 averages seven pollsters' latest polls of the line-up most of them test. It puts Marine Le Pen on 34.1 per cent in round one, Édouard Philippe on 17.6 and Mélenchon on 16.8, with Raphaël Glucksmann on 11.0 and Bruno Retailleau on 7.4. The gap for second place is under a point. Our presidential page estimates that polls this far out have moved by about 5.2 points per candidate before polling day.

Le Pen's place in the race was largely settled in July. The Paris Court of Appeal upheld her conviction for misusing European Parliament funds but cut her ban from office, which Jurist reported left her free to stand; she said she would run and appeal to the Court of Cassation. Toluna Harris Interactive's poll for M6 and RTL in late September, reported by Euronews, put her on 35 to 36 per cent.

Why it is not an artefact

The first check is the overlap in our definition. The month-out average includes the final polls, because its window runs to the end of the campaign. That overlap shrinks the measured slide rather than creating one. When we used a window that ends a month before polling day and shares almost no polls with the final set, 37 of 55 parties that fell two points or more still finished below their final polls.

The second check concerns noise. Because the final polls appear in both comparisons, a set of finals that happens to read low makes a party look as if it fell late and then makes the result look like a recovery. Justin Wolfers, writing for the Brookings Institution, made this point in 2010 about tests of poll momentum: shared measurement error pushes such tests towards finding reversals. Our test is tilted against the pattern, and the pattern still shows.

The third check is selection. Dropping parties whose result was below 5 per cent would have removed the cases that fell furthest, so we filtered on the final polls instead. That brings in Pécresse. On the result filter the figure is 23 of 28.

The fourth check is whether these are simply parties that polls always overstate. Compared with each party's own polling miss at its other elections, 20 of 27 sliders still did worse than usual. The count also survives other cuts. With a 1.5-point threshold it is 31 of 42; at three points, 12 of 16. Since 2000 it is 18 of 21, with an average miss of 1.2 points. Leaving out Canada, the country with most cases, gives 17 of 20; leaving out Britain gives 19 of 22.

The exceptions and the limits

Five sliders recovered. Labour in 1983 fell from 30.6 to 25.6 in the final polls and then won 28.3. The Canadian Liberals in 2006 dropped to 28.0 in the finals and won 30.2. Geert Wilders's PVV in 2025 slid from 19.5 to 16.2 and won 16.7. A late slide is a signal, not a rule.

The sample is also small. Twenty-nine cases from 65 elections is enough to show a pattern well beyond chance, but not enough to say how large the extra fall will be in any one race. The median was about one point. Nine cases come from Canada and seven from Britain, so two national polling cultures carry weight. Nothing in our data says why the pattern exists, and we do not offer a cause.

It also cuts against an old finding. Nate Silver, writing in 2010 about US Senate races since 1998, found that a candidate's gains in one period of polling did not predict further gains, a result the Democratic Strategist summarised as no reason to expect trends to continue. Sam Wang of the Princeton Election Consortium has argued that late polling errors run in an unpredictable direction. Our data agree for parties that are rising. For parties that are falling in the last month, the record runs the other way.

What to watch next spring

The final month of the French campaign starts in mid-March. If any of the leading candidates loses two points or more between then and the final polls, the record says to expect the slide to go further on 18 April. With Philippe and Mélenchon under a point apart, a candidate falling late could miss the run-off by more than the final polls suggest, as Jospin did in 2002. Our projection does not adjust for this, and a single poll moving a point or two is normal scatter. A slide across most pollsters is the signal worth reading.