Forecasting long memory left–right political orientations

Abstract

This paper considers out-of-sample forecasting of left–right political orientations of party affiliates in the Netherlands, using weekly data from 973 independent national Dutch surveys conducted between 1978 and 1996. The orientations of left-wing and right-wing party affiliates tend to converge over time in the sense that the differences between the average positions tend to decline. The left–right series also reveal long-memory properties in the sense that shocks appear to be highly persistent. We develop forecasting models that account for these data features and we derive the relevant forecast intervals.

Introduction

Just as liberal–conservative is the key concept in US electoral research, left–right is central to studies of political party competition in Western Europe. Being the best single explanatory variable for vote intention and political party choice, almost every political survey conducted in Western Europe in the past several decades has incorporated the left–right political orientation scale into its questionnaire. However, because there seems to be a consensus among political scientists that left–right orientations are permanent and invariable, possible changes in left–right orientations among West European mass publics have drawn little scholarly notice. This paper draws upon 973 independent national surveys conducted between 1978 and 1996 to track trends in the aggregate left–right political orientations of party affiliates in the Netherlands. More specifically, we examine whether the average left–right orientations of the party groupings tend to converge to each other and whether these changes are of a long memory nature. We subsequently formulate forecasting models that account for these data features.

The novelty of this paper is not in the application of long memory models to political data. Some previous studies have applied the autoregressive fractionally integrated moving average (ARFIMA) model to analyze the long-memory properties of the outcomes of political polls in the US and the UK. Box-Steffensmeier and Smith (1996) have estimated ARFIMA models for the percentages of Republican and Democratic identifiers in 160 quarterly polls from 1953 to 1992 in the US and Byers et al. (1996) have estimated these models for (logistic transforms of) Conservative and Labor support in 417 monthly polls from 1960 to 1995 in the UK. Both studies find that aggregate series of partisanship are pure fractionally integrated processes. In our study, however, the estimates for the MA parameters turn out to be significant and we determine the significance of the drift in the data, an issue which previous studies did not address. Additionally, we experiment with different deterministic trend functions, including a logistic trend. Our paper is also novel – next to the large number of time points we use – in that it clearly demonstrates the crucial part played by the fractional differencing parameter in long-range forecasting.

The paper is organized as follows. Section 2examines the data and some of their properties. Section 3incorporates the convergence and long-memory properties of the data into forecasting models and Section 4reports the empirical results. It derives forecast intervals and investigates whether the average left–right orientations will reach the state of convergence in the near future. The paper concludes with some remarks in Section 5.