Forecasts From Nonrandom Samples

Abstract

The 1990s was not the best of decades for electoral polls, with striking errors occurring in, among others, the British, French, and Spanish elections, including election night, when errors are more evident. This article proposes a model for predicting final election outcomes based on the consistency that polling stations show between elections. Using both past and incoming polling station vote proportions, the model produces continuously revised predictions. The method is validated predicting the 1995 Corts Valencianes (Valencia regional parliament) elections and displaying the real-time experience of the 1999 Corts Valencianes election night. The case study is completed by demonstrating the technique's efficacy in three additional elections. The results confirm that the procedure generates quick, highly reliable, and accurate forecasts. In fact, only a few minutes after starting the scrutiny, the proposal permits one to approximate the final results with great precision, even with only a small percentage of votes polled. The great flexibility of the procedure makes it possible to use the method under a wide variety of circumstances and electoral systems. Furthermore, this procedure has additional advantages, including robustness and lower cost, over other methods which can also be implemented during election night with the objective of forecasting final outcomes, like exit polls or quick counts of a meaningful sample of polling stations.