A Fundamentals-Based Forecast of the 2018 Italian Elections

A comparison of our predictions with the mean of surveys conducted in Italy during the first week of February 2018.

The table above presents preliminary results from our project on 2018 Italian Parliamentary Elections. The aim of the project is to come up with a fundamentals-based forecast of the vote shares and resulting seats' allocation for the main parties and coalitions running on March 4th. By fundamentals-based, we mean that our prediction is grounded on what Italian voters are expected to do based on the recent dynamics of the two most debated issues: the migration crisis and the performance of the economy. To come up with our estimates, we build a bread-and-peace type model. In our framework, the vote shares of the main parties at the level of 7543 municipalities are the result of two fundamental elements: a baseline expected vote share plus a variation around the baseline, depending on a relatively small set of socioeconomic factors.

The mean of our predictions for the national vote share of the center-right  coalition, plotted against the mean of surveys in each month going from January 2017 to January 2018.

The baseline is jointly determined by the past vote share earned by party p in municipality i and a province fixed effect, aimed at capturing persistent geographical ideologies across the Italian territory. It is important to note that, upon observing the electoral dynamics in Italy during the last 25 years, we have decided to use electoral results in t-2, that is, two races before the one we are predicting, rather than in t-1. Therefore, in the case of 2018, the relevant shares included in the model are those obtained by parties in the elections of 2008. Once this starting point is set, our model predicts that a deviation will take place, based on how the state of the local economy and the migration phenomenon have shaped voters' preferences throughout the term. Specifically, in order to gauge the effect of possible occupational shocks  at the municipal level, we source information about the share of enterprises and jobs in the municipality that used to accrue to the manufatcuring sector as of 2011. Then, we interact these measures of the importance of manufacturing with the relative change in the unemployment rate at the provincial level during the last term preceding the election. This allows us to compute a measure of vulnerability to occupational shocks that varies across both terms and municipalities. With a similar spirit, and in order to account for the preferences of younger voters, we interact the share of people in the 18-29 age bracket at the municipal level with the provincial change in the inactivity rate for people of that age. As far as immigration is concerned, our proxy for the magnitude of this phenomenon is the relative variation of the share of foreign-born residents in the municipal population, again computed on a term basis. The equation is completed by controlling for the natural logarithm of the municipal population, and estimated via seemingly unrelated regression (Zellner 1962). This technique has indeed proved to be particularly suitable for multiparty data (Tomz et al. 2002), as it estimates all the equations for each party simultaneously. As a result, it takes into account that if party j is getting x% more votes, then they must be either sourced from other parties' electorates or from people who did not turn out in the previous election.

The mean of our predictions for the national vote share of the Democratic Party (center-left), plotted against the mean of surveys in each month going from January 2017 to January 2018.

Rather than coming up with a single prediction of the electoral performances of parties in each municipality, we follow King et al. (2000) and implement 1000 simulations of the predicted vote shares determined by our model. This maximizes the use of the information at our disposal, while exploiting the intrinsic uncertainty of our framework to produce estimates of Bayesian credibility intervals at the party-municipality level. The process leads to the results presented in the table above and in the plots. In a subsequent stage, by taking advantage of the extremely fine-grained nature of our data, we apply the electoral formula currently in force and aggregate votes at the level of electoral districts. This is done by weighting each municipality by the predicted number of people that will turn out, thus transforming vote shares into actual numbers of votes, and summing them up across all the municipalities in district j. The total number of votes provides us with a ranking of parties and coalitions in each of 231 single-member district. Following the electoral rules, we assign the seat to the first-ranked party or coalition of parties. The remaining seats are instead distributed by looking at the shares that parties are expected to gain at the national level. In fact, even if 387 multi-member districts have been designed, electoral prescriptions establish that they will matter only to determine where parties are going to earn their seats, but not how many of such seats they will be eventually entitled to get.

The mean of our predictions for the national vote share of the Five-Star Movement, plotted against the mean of surveys in each month going from January 2017 to January 2018.

In order to be used to predict the results of March 2018, the model had first been trained with an out-of-sample prediction of the results of the last Italian parliamentary election, which took place on February 2013. This means that we have adapted it to that election, and estimated the coefficients of each of our predictors on a pool of 5753 randomly drawn municipalities, which constituted our training set. After getting an estimate of how our variables had contributed to determine actual vote shares in those municipalities, we have used the coefficients to predict the results in the remaining 2005 units of observation, which formed our test set. The results, displayed in the table below, are quite encouraging.

The performance of our model in forecasting out-of-sample results for 2013 in 2005 Italian municipalities.

                            

Namely, the percentage-point differences between predicted and actual vote shares always lie in the 0.24-2.8 range. Moreover, Root Mean Squared Errors at the level of individual municipalities are remakably low, with the partial exception of those relative to the performance of the Five-Star Movement. However, this is probably due to the fact that we cannot include a baseline vote share for this party, as it was not on the political arena on 2006, which constitutes our reference election year for the prediction at stake. An identical consideration applies to the predictions for 2018, which use 2008 as the baseline election. In spite of these issues, however, the ranking of coalitions implied by our exercise mirrors the one depicted by surveys. Furthermore, the discrepancy between our forecast and the vote share computed by the most recent polls is quite limited. Notably, as one can easily see from the plots, most of the gap between our figures and those of surveys has opened up only in recent months. This is by no means surprising. In fact, as elections approach, survey interviews can obviously measure preferences for otherwise unobservable factors such as leaders' valence, candidates' qualities, or the effect of shocks that have hit one or more parties. For instance, it is extremely likely that the marked decrease in the vote share for the Democratic Party predicted by polls conducted in the last 6 months is reflecting the split that occurred within it, and eventually led to the formation of the leftist party Liberi e Uguali. Nonetheless, it is interesting to see how our framework, without resorting to any interview, is capable of producing credible estimates of a future electoral race based on a few, carefully chosen predictors.

About authors

Massimo Pulejo is a Pre-Doctoral Fellow at Bocconi University, Department of Policy Analysis and Public Management.
Piero Stanig is Assistant Professor of Political Science at Bocconi University, Department of Policy Analysis and Public Management, and a fellow of the Carlo F. Dondena Research Center. 

Thanks to Giovanni Da Fre' for excellent research assistance.