Measuring Italian polling error, 2008–2022
Abstract. We measure the error of the final polls at the Camera elections from 2008 to 2022, all taken at least fifteen days before the vote because of the publication ban: which polls count, how their sources are graded, the error list by list and on the gap between the two largest lists, each pollster's record and the long view. The M5S was understated at every election it contested, by every pollster in 2013 and 2018; the PD was overstated at three of four. Earlier final polls give coalition totals only.
Summary
This paper sets out how we measure polling error in Italy: which polls count as final polls, how the fifteen-day ban on publishing polls shapes them, how their sources are graded, how the error is defined, and what the record shows. At the four Camera elections from 2008 to 2022 the average of the final polls missed each list by between 1.5 and 3.2 points, more than in any other country on the site. One list stands out. The Five Star Movement (M5S) was understated at all three elections it contested, by 10.4 points in 2013, 4.8 in 2018 and 2.0 in 2022, and in 2013 and 2018 by every pollster. The PD was overstated at three of the four elections. These two misses have kept their direction, and they are why, unlike in Germany and Spain, the correction passes its test in Italy (Reading Italian polls: the correction and its test).
The record has three uses on the site: its size sets the width of the ranges (Simulating Italian polling error: where the chances come from); its direction sets the correction (Reading Italian polls: the correction and its test); and its pollster-by-pollster detail feeds the ratings (Rating Italian pollsters: accuracy, the lead and openness).
What has changed since this paper was written
Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.
- Only elections from 2000 shape the forecast (28 September). Earlier elections stay in the long view and on the history pages but no longer feed any test, correction or error size. In Italy the tests were already on elections since 2001, so nothing changed.
- When the polls missed (2 November): Grillo's tsunami: how Italy's 2013 polls missed the Five Star surge by ten points.
- The track record across countries (28 September). Final polls at 133 elections in nine countries, measured the same way, are at /accuracy/elections.
What is measured, and against what
The quantity that decides the proportional seats is each list's national share of the valid vote, and that is what Italian voting-intention polls report. We measure against the national list shares of the Camera, for Italy excluding the Valle d'Aosta and the constituency abroad, which is the base of the proportional count. Results come from the Ministry of the Interior's figures (The Italian data: results, maps, statistics, deputies and polls).
Lists that changed name or merged are measured as lineages:
- PD: the PD from 2008 (and the Ulivo list of 2006, although 2006 cannot be scored).
- FI: the Popolo della Libertà in 2008 and 2013, Forza Italia since.
- Lega: the Lega Nord until 2018, the Lega since.
- AVS: Sinistra Ecologia Libertà in 2013, Liberi e Uguali in 2018, Alleanza Verdi e Sinistra in 2022.
- FdI and the M5S as themselves.
The lists measured at each election are those in the result and in at least half of that election's final polls. The six above are the ones the correction uses. Everything else is not scored.
Definitions
Let \(v_{p,t}\) be list \(p\)'s share at election \(t\) and \(y_{j,p,t}\) its share in final poll \(j\). The error of a poll is \(e_{j,p,t} = y_{j,p,t} - v_{p,t}\): poll minus result, so a positive error means the polls overstated the list. The industry error is the error of the simple average of the final polls,
with each pollster counting once. The average miss per list at an election is
and each list's record is summarised by its mean error (direction) and root mean square (size):
The gap error on the two lists that finished first and second, \(a\) and \(b\), is \(e^{\text{gap}}_t = e_{a,t} - e_{b,t}\); a positive value means the polls put the winner too far ahead.
The evidence: final polls under a fifteen-day ban
What counts as a final poll
Article 8 of legge 28/2000 forbids making public the results of polls on voting intentions in the fifteen days before the vote, even for polls taken before the ban began. A pollster's final poll is therefore its last poll whose fieldwork ended at least fifteen days before polling day, and no more than 75 days before. Polls that Wikipedia lists inside the ban (in 2013, polls published only after the vote) are excluded. One poll per pollster per election is kept.
The ban makes Italian final polls early:
| Election | Final polls | Median days before the vote | Closest to the vote | Furthest |
|---|---|---|---|---|
| 2001 | 5 | 20 | 17 | 40 |
| 2006 | 6 | 19 | 17 | 33 |
| 2008 | 15 | 28 | 17 | 62 |
| 2013 | 18 | 19 | 16 | 58 |
| 2018 | 17 | 19 | 16 | 66 |
| 2022 | 21 | 18 | 16 | 25 |
The typical final poll finished fieldwork eighteen or nineteen days before the vote in 2013, 2018 and 2022, and 28 days before in 2008. A Spanish final poll is about ten days old on polling day and a British one a day or two. Everything that moves in the last two and a half weeks of an Italian campaign is invisible to the polls and counts, in our record, as polling error.
The elections
The database holds final polls for seven Camera elections, but only four can be scored list by list. In 1996, 2001 and 2006 the final polls we could source report coalition totals only (the Ulivo, the Casa delle Libertà and so on), and the lists inside them cannot be separated. That leaves 82 final polls at the elections of 2008 (15), 2013 (18), 2018 (17) and 2022 (21). No pollster has a final poll at more than four of them.
Grading the source
Every figure carries a source grade. Grade A is the pollster's own release or the official register of published polls; grade B is the article of the outlet that commissioned the poll; grade C is a secondary transcription, here Wikipedia's poll tables. The official register (sondaggipoliticoelettorali.it), where every published poll's full document must be filed, refused every connection from the network our data was collected from, and most pollsters' old releases are gone. As a result almost every final poll is grade C: one 2022 poll could be confirmed from the commissioning outlet. The Italian data: results, maps, statistics, deputies and polls describes the checks.
A worked example: 2022
The 2022 election, held on 25 September after the Draghi government fell, had 21 final polls:
| Pollster | Fieldwork end | FdI | PD | M5S | Lega | FI | AVS | Poll's miss per list |
|---|---|---|---|---|---|---|---|---|
| Izi | 8 Sept | 25 | 21.4 | 15.4 | 10 | 8.6 | 4.2 | 0.94 |
| Hokuto | 7 Sept | 26.8 | 20.2 | 14.5 | 11 | 7 | 3.5 | 1.06 |
| Ipsos | 6 Sept | 25.1 | 20.5 | 14.5 | 12.5 | 8 | 3.4 | 1.22 |
| Cluster17 | 8 Sept | 24.4 | 20.8 | 14.1 | 11.3 | 8.4 | 3.6 | 1.25 |
| Index Research | 7 Sept | 26.5 | 21.1 | 14 | 11 | 7 | 4 | 1.28 |
| EMG | 7 Sept | 24.5 | 19.5 | 13 | 12 | 8 | 3 | 1.39 |
| Quorum / YouTrend | 9 Sept | 25.3 | 21.2 | 13.8 | 12.9 | 7.9 | 3.4 | 1.51 |
| CISE | 5 Sept | 23 | 21.4 | 16.6 | 9.6 | 8 | 5.9 | 1.62 |
| Ixè | 2 Sept | 22.7 | 21.7 | 13.8 | 10.9 | 7.9 | 3.5 | 1.67 |
| Demos & Pi | 7 Sept | 24.6 | 22.4 | 13.8 | 12 | 7.7 | 3.4 | 1.71 |
| Noto Sondaggi | 7 Sept | 24.3 | 19.5 | 13.5 | 13 | 6.9 | 2.7 | 1.74 |
| BiDiMedia | 9 Sept | 25.8 | 22.3 | 13.1 | 11.8 | 6.7 | 4 | 1.76 |
| Euromedia Research | 7 Sept | 24.7 | 21.8 | 13 | 11.8 | 7.2 | 3.8 | 1.76 |
| SWG | 9 Sept | 27 | 20.4 | 12 | 12.1 | 6.7 | 4 | 1.82 |
| GPF | 4 Sept | 24.4 | 22.6 | 13.4 | 12.7 | 8 | 3.9 | 1.91 |
| Tecnè | 8 Sept | 25.4 | 21.6 | 12.2 | 11.4 | 10.2 | 3.2 | 1.92 |
| GDC | 7 Sept | 24.4 | 21.5 | 12.3 | 12.5 | 7.3 | 3.2 | 2.02 |
| Termometro Politico | 8 Sept | 25.2 | 22.2 | 13.3 | 13.3 | 7.2 | 2.6 | 2.09 |
| Demopolis | 7 Sept | 25 | 22.4 | 12.8 | 13.5 | 6.9 | 3.5 | 2.17 |
| Piepoli | 9 Sept | 23.5 | 21.5 | 11.5 | 12 | 9.5 | 4 | 2.31 |
| Lab21.01 | 9 Sept | 23.3 | 24.7 | 12 | 15.7 | 6 | 2.6 | 3.64 |
| Average | 24.80 | 21.46 | 13.46 | 12.05 | 7.67 | 3.59 | ||
| Result | 25.99 | 19.04 | 15.43 | 8.79 | 8.11 | 3.64 | ||
| Error | −1.2 | +2.4 | −2.0 | +3.3 | −0.4 | 0.0 |
Table: Final polls and result, Camera 2022, national list shares, points. Rows ordered by each poll's own miss.
The average missed by 1.55 points per list. Every final poll had the Lega too high, by between 0.8 and 6.9 points, and every one but CISE had the M5S too low. Every poll had the PD too high, by between 0.5 and 5.7 points. The polls had FdI 3.3 points ahead of the PD; it won by 6.9, a gap error of \(-3.6\). Every poll was taken at least sixteen days before the vote.
The record, 2008–2022
| Party | 2001 | 2006 | 2008 | 2013 | 2018 | 2022 | Average |
|---|---|---|---|---|---|---|---|
| PD (Ulivo 2006) | −1.4 | +1.0 | −0.2 | +4.7 | +3.9 | +2.4 | +1.7 |
| FI (PdL 2008–13) | −1.7 | −2.2 | +1.5 | −1.8 | +2.4 | −0.4 | −0.4 |
| Lega | +0.5 | −0.4 | −3.0 | +1.1 | −3.8 | +3.3 | −0.4 |
| FdI | – | – | – | −0.1 | +0.4 | −1.2 | −0.3 |
| M5S | – | – | – | −10.4 | −4.8 | −2.0 | −5.7 |
| AVS (SEL, LeU) | – | – | – | +0.9 | +2.3 | 0.0 | +1.0 |
Four points stand out.
The M5S has always been understated. By 10.4 points in 2013, when every final poll had it between 6.8 and 13.1 points too low; by 4.8 in 2018, again by every poll; and by 2.0 in 2022, by all but one. The miss has shrunk as the party has aged, but not changed sign. Several explanations have been offered: late movement during the ban, supporters who distrust pollsters and are harder to reach, and a vote concentrated in the South, where turnout is harder to predict. Our data cannot separate them.
The PD has been overstated since 2013. By 4.7 points in 2013, 3.9 in 2018 and 2.4 in 2022, each time by every pollster. In 2008 it was polled almost exactly.
The Lega changes sign. It was understated by 3.0 points in 2008 and 3.8 in 2018, and overstated by 1.1 in 2013 and 3.3 in 2022. A correction learns almost nothing from it.
2013 was the worst. The average missed by 3.2 points per list. Every poll had the PD well ahead; the M5S finished level with it as a list.
Using equation (3):
| List | Elections | Mean error | RMS error | Share of RMS² that is bias |
|---|---|---|---|---|
| M5S | 3 | −5.73 | 6.74 | 73% |
| PD | 4 | +2.70 | 3.28 | 68% |
| Lega | 4 | −0.60 | 2.95 | 4% |
| FI | 4 | +0.43 | 1.68 | 6% |
| AVS | 3 | +1.07 | 1.41 | 56% |
| FdI | 3 | −0.30 | 0.72 | 17% |
Table: Summary of industry errors, 2008–2022, points. The last column is \(\bar{e}_p^2 / \text{RMS}_p^2\).
This is the opposite of the German and Spanish pattern. For the M5S and the PD most of the error is steady; for the Lega and FI almost none is. A correction that takes off half the mean will help most where the steady part is large, and do little harm where it is small, because there the mean is close to zero. That is what the test finds (Reading Italian polls: the correction and its test).
Shared and pollster-specific error
The average squared poll error splits into the square of the industry error and the spread of the polls around their average (Measuring British polling error in England, 2015–2024). For the large Italian misses the shared part is very high: 98% for the M5S in 2013, 95% in 2018; 90% for the PD in 2013 and 94% in 2018; 94% for the Lega in 2018 and 87% in 2022. The pollsters missed together. No choice among them would have avoided it, and averaging more polls does not reduce it.
The gap between the two largest lists
| Election | Top two lists | Real gap | Polls' gap error | Worst poll |
|---|---|---|---|---|
| 2008 | PdL, PD | 4.2 | +1.6 | +7.3 |
| 2013 | M5S, PD | 0.1 | −15.1 | −20.4 |
| 2018 | M5S, PD | 13.9 | −8.8 | −13.0 |
| 2022 | FdI, PD | 6.9 | −3.6 | −8.3 |
Table: Gap between the two largest lists, points.
The M5S's record shows in the gap. In 2013 the polls had the PD about fifteen points ahead of the M5S; they finished level. In 2018 the polls had the M5S about five points ahead of the PD; it won by fourteen. In Italy the lead matters less than in Spain or Germany, because seats go to coalitions, but the same misses move the coalition totals.
Pollster by pollster
The table lists pollsters with three or more final polls; the track record page lists all of them.
| Pollster | Elections | Average miss per list |
|---|---|---|
| IPR Marketing | 3 | 2.28 |
| Ipsos | 4 | 2.30 |
| Euromedia Research | 4 | 2.33 |
| SWG | 3 | 2.39 |
| Demopolis | 4 | 2.42 |
| EMG | 3 | 2.43 |
| Demos & Pi | 4 | 2.52 |
| Termometro Politico | 3 | 2.63 |
| Quorum / YouTrend | 3 | 2.63 |
| Tecnè | 3 | 2.72 |
| Lorien Consulting | 3 | 2.84 |
| BiDiMedia | 3 | 3.02 |
| Piepoli | 4 | 3.30 |
Table: Average miss per list of each pollster's final polls, 2008–2022, points.
The spread between pollsters is narrow, from 2.3 to 3.3, next to the spread between elections, from 1.5 to 3.2. Every pollster with a 2013 final poll carries that election's miss. The ratings (Rating Italian pollsters: accuracy, the lead and openness) use these records, pulled towards a poor Italian final poll.
How the record is used
Our rule takes off half the average miss and keeps the correction only if it would have improved past forecasts, each election predicted without knowing its result. In Italy it does, by a small margin: 1.82 points per list against 1.92 (Reading Italian polls: the correction and its test). The record also sets the size of the simulation's error for each list (Simulating Italian polling error: where the chances come from): the M5S's typical miss of more than six points makes its range the widest of any list.
The long view
| Election | Final polls | Average miss per party | Biggest miss |
|---|---|---|---|
| 2001 | 5 | 1.19 | FI (PdL 2008–13) −1.7 |
| 2006 | 6 | 1.19 | FI (PdL 2008–13) −2.2 |
| 2008 | 15 | 1.52 | Lega −3.0 |
| 2013 | 18 | 3.14 | M5S −10.4 |
| 2018 | 17 | 2.95 | M5S −4.8 |
| 2022 | 21 | 1.55 | Lega +3.3 |
Before 2008 the record is thin. The only party-level final poll we could source before 2008 is one Abacus poll of April 2001, whose figures are coalition totals and the lists that stood alone. For 1996 and 2006 the final polls give coalition totals only, and for 1994 we found none. We collected Camera results back to 1948 for context, but the long view of polling error in Italy starts in practice in 2008.
Limits
Four elections. Every figure here rests on four numbers per list, three for the M5S, FdI and AVS. A fifth election in which the M5S was overstated would change its mean sharply.
Secondary sources. Almost every final poll is a Wikipedia transcription. The official register would give the full documents, and could not be reached.
The ban. Final polls are two and a half weeks old on polling day. The record mixes polling error with late movement, and the two cannot be separated.
Lineages. The PdL and FI, the Lega Nord and the Lega, SEL, LeU and AVS are treated as continuous. Each merger and split changes what the polls measure.
Final polls only. These errors say nothing directly about a poll taken a year out; that is From today to polling day in Italy.
References
Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.
Legge 22 febbraio 2000, n. 28, article 8.
Shirani-Mehr, H., Rothschild, D., Goel, S. and Gelman, A. (2018). Disentangling bias and variance in election polls. Journal of the American Statistical Association, 113(522), 607–614.
How to cite
Polling Forecast (2026). Measuring Italian polling error, 2008–2022. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/italy-polling-error. Live figures change as new polls arrive; the date above is when the text was last revised.



