Abstract

This study examines electoral forecasting in volatile party systems, focusing on factors contributing to deviations between poll predictions and actual election outcomes. Using Italy as a case study, it identifies biases in polling data and proposes a method to enhance estimator accuracy in a context of stable institutions and volatile electoral dynamics. Data from three Italian general elections are analyzed to evaluate discrepancies between pre-electoral polls and results, assessing key factors such as timing of data collection, survey methodology, sample size, and party system fragmentation. Employing a Bayesian inference process via a Markov chain Monte Carlo (MCMC) adaptive Metropolis-Hastings (MH) algorithm, the study demonstrates that pre-electoral estimates can be significantly improved using the Two-Stage Model (TSM). By consistently outperforming traditional poll predictions, the TSM offers a robust framework for addressing polling biases. These findings advance political forecasting by improving accuracy in both consolidated democracies and volatile electoral contexts, while emphasizing the need for future research on dynamic polling methods and fundamentals-based models.