President Barack Obama and Prime Minister Matteo Renzi of Italy participate in a press conference in the East Room of the White House, April 17, 2015. (Official White House Photo by Pete Souza)

Can Renzi win the Constitutional referendum in Italy?

“I’m rooting for [Matteo’s] success” said US President Barack Obama upon receiving Italian Prime Minister Matteo Renzi at the White House for his final state dinner on October 18. “I won’t weigh in on the referendum, but the reforms Matteo is initiating, certainly on the economic side, are the right ones” he further added.

The Italian referendum, which largely aims at transforming the Senate in a consultative body and re-organising the distribution of competences between the central and regional governments, comes however at a difficult time. Voters are being called to accept or reject far-reaching changes to the Constitution in the NO-year of referenda across the world. Brexit for Britain, Farc in Colombia, EU-Ukraine partnership in the Netherlands all resulted in a blow to the ruling elites. The question is whether Italy will confirm the rule or turn out to be an exception.

Polls are not looking extremely favourable to the reformist Italian Prime Minister. Eight years ago, Obama himself campaigned on a platform of Change, but only “change you can believe in”. The ultimate question is then: with one month to go, can Renzi still turn tables and win the Constitutional referendum?

Model predicted results if the referendum were held on 3 November 2016

To answer this question, I built a basic polynomial probabilistic model based on all available polls (see Methodological note for more details). Of course, polls can be inaccurate, as the Brexit case proves, or can be quickly turned around by last minute news, as the FBI-case on Clinton shows in America. As such, results should be always taken with a pinch of salt.

Following a long stream of polls suggesting Renzi’s Yes campaign is lagging behind, it comes as no surprise that the model forecasts with a probability in the high 90% that if the referendum were held today, Renzi would lose — though the gap between No (51.9%) and Yes (48.1%) is far from huge. As such, even small last minute changes in turnout could easily tip the balance, in one direction or another.

Share of YES in all polls between 1 September and today, and model tracking prediction with 95% confidence interval

However, the more interesting question is: what will happen on December 4th? The model can be used for out-of-sample extrapolations, to get a sense of what the referendum result could potentially be, based on current polling trends.

The YES camp used to be far ahead in the polls during the summer. At that point, all oppositions to Mr Renzi, ranging from the far-left to the far-right, have managed to converge and present a united NO front. Since then, the YES camp has started losing ground and ditched below the crucial 50% threshold.

However, the model captures the fact that this decreasing trend was interrupted around mid-October, and there are some preliminary signs of a recovery. While it’s hard to determine exactly what has led to this change in tide, it’s interesting to notice how this comes on the back of an intensified media presence of the Prime Minister, characterised by two important TV debates (22 and 30 September) and a flurry of other appearances, including radio.

At the current stage, the model assigns a 59% probability to Mr Renzi eventually prevailing. However, the predicted margin of victory is extremely slim (50.9% to 49.1%) and final results will crucially rest on how polls will evolve over the next few days. If the Prime Minister fails to reinforce this new upward trend, this winning probability will inevitably recede.

Model-based winning probability for December 4th

Over the coming days, I will continue re-running the model as fresh polls are released and updating the predictions. Stay tuned!


I would like to thank Michele Peruzzi for statistical advice and Luca de’ Angelis for comments on an earlier draft of this article.

Disclaimer: All views, errors, and omissions are my own and not to be associated to the institutions I am affiliated with.

Methodological note

The model is based on a polynomial fit, analytically weighed by the inverse of the poll’s sampling error. Various specifications were tested, and selected based on the highest Adjusted-R2.

The data used are all available polls from 1 September to today. Only polls providing a margin of error and exact Yes/No share were utilised. Furthermore, an outlier analysis was carried out and the most extreme outlier was excluded.

Obtained from