Applying Machine Learning Methods to Forecast Japanese Elections

• In Event: Forecasting Political Outcomes with Different Methods

Abstract

This paper applies machine learning methods, namely, ensemble learning and support vector regression (SVR), to forecast Japanese lower house elections. We used these two methods to reduce problems of small sample size, which are inherent in any single-country forecasting model, and, at the same time, to realize more accurate predictions through sophisticated non-linear forecasting models. By applying these two machine learning methods, we developed several non-linear forecasting models. We then compared our forecasting results with those generated by Lewis-Beck and Tien (2012), a pioneering model on this topic, using the same data and variables. We found that, even without tuning, virtually all of our ensemble-learning-based forecasting models exceeded Lewis-Beck and Tien’s model in accuracy and provided better explanations for variance. Further, one of our models utilizing eXtreme Gradient Boosting was shown to substantially enhance generalization performance with appropriate tunings of hyper parameters. We also analyzed, by calculating feature importance, the variables that played a significant role in determining each electoral outcome.

Few attempts have been made to forecast Japanese elections in the past. One of the reasons for such underdevelopment is a serious, small-sample problem that is inherent in electoral forecasting of any single country. Only 26 lower house elections have been held in Japan following the country’s full democratization at the end of World War II. Such sample-size constraints hang especially heavily on constructing sophisticated, non-linear forecasting models. The pioneering work of Lewis-Beck and Tien thus used a very simple linear model to forecast the percentage of seats to be won in lower house election by the Liberal Democratic Party (LDP), which has dominated Japan’s postwar, political landscape. The larger sample sizes of cross- national electoral forecasting models, on the other hand, have recently enabled the development of non-linear forecasting models utilizing machine learning techniques (e.g., Kennedy et al. 2017). Such cross-national forecasting models obviously cannot, however, account for country-specific institutional and behavioral aspects of elections that might substantially enhance the models’ accuracy.

We developed non-linear forecasting models for Japanese elections by using ensemble learning and support vector regression (SVR) methods. These two machine learning methods are known to be effective when sample size is small. We used the techniques of these two methods to develop non-linear forecasting models while minimizing the small sample problem. More specifically, as for parallel ensemble methods, we used Bagging and Random Forest; as for sequential ensemble methods, we used AdaBoost and Gradient Boosting; and as for SVR, we constructed a model using the radial basis function and sigmoid function as the kernel.

To construct and evaluate our models, we divided the dataset into a training dataset and testing dataset, using the former to adjust hyper-parameters by cross validation and then evaluating generalization performance using the latter. As for indicators to evaluate the models, we used maximum absolute error and explained variance. All our six models obtained better scores for the two indicators compared to Lewis-Beck and Tien’s model. Among them, the model using eXtreme Gradient Boosting performed the best; it reduced maximum absolute error by 2.4 points and increased explained variance by 2.67 points. We calculated the feature importance of each variable we used in our models and found that all were similarly important. We also examined what new variables may be added to our (and Lewis-Beck and Tien’s) models to increase predictability.

Overall, our non-linear machine-learning approach to Japanese electoral forecasting showed promising results. Ensemble learning and SVR methods reduced the problem of small sample size. All our models exceeded Lewis-Beck and Tien’s pioneering model in accuracy and explained variance. We further showed that, with appropriate tuning for hyper parameters, the predictability of our model using eXtreme Gradient Boosting could be substantially enhanced. Our approach allows researchers to incorporate country-specific institutional and behavioral aspects into non-linear forecasting models, thereby alleviating, to a certain extent, the problem of small sample size. We believe our approach can also be applied to other country-specific forecasting models and substantially enhance their predictability.

Authors

• Sota Kato, International University of Japan

• Hirokazu Shimauchi, The Tokyo Foundation for Policy Research

• Budrul Ahsan, Philips Japan

• Takafumi Nakanishi, Musashino University