Predicting Australian federal electoral seats with machine learning

I expand the international literature on election forecasting with the first application of machine learning (ML) in the Australian context. I apply these models to five elections from 2010 to 2022 and compare them with the dominant forecasting tool in Australia, the Mackerras pendulum. I evaluate these models’ accuracy in predicting the winning party for each electoral seat and estimating the total number of seats won by each party. Pendulum forecasts corrected with an extra trees model that incorporates state effects, seat-level unemployment rate, and vote share history predict up to 15 additional seats correctly six to three months before each election. The traditional pendulum is increasingly strained by polling errors and a larger crossbench. New modeling techniques will only emerge through experimentation. This study demonstrates the potential for ML-based election forecasting in Australia and provides a starting point for further efforts to surpass the pendulum.