Do Econometric Models Provide More Accurate Forecasts When They are More Conservative? A Test of Political Economy Models for Forecasting Elections

Abstract

The assumptions of multiple regression analysis are not met in many practical forecasting situations and, as a result, regression models are insufficiently conservative. We tested the effect on forecast accuracy of applying three evidence-based forecasting guidelines to 18 political economy models for forecasting elections in nine countries, all of which were originally estimated using multiple regression analysis.

The guidelines direct modelers to account for uncertainty of econometric model forecasts by:

(1) modifying estimates of the strength of variable effects,

(2) combining forecasts from diverse models, and

(3) taking account of all variables that are known to be important.

Out-of-sample forecast accuracy was compared with the accuracy of forecasts from the originally published econometric models representing typical practice. While damping the estimated variable weights did not improve accuracy, equalizing them reduced error compared the original model forecasts by 10%. Combining forecasts from models for US (N=8) and Australian (N=2) elections reduced error by 25% on average. Including more causal knowledge, by using all unique variables from the different models in equal-weights index models, reduced error on average 26%.

Keywords: combining forecasts, damping, equalizing, elections, golden rule of forecasting, index method, shrinkage