We explore the aspects that affect the forecasts of Large Language Models (LLMs) for an electoral outcome in Victoria, Australia. In particular, the popularity of One Nation has created considerable uncertainty about the upcoming state election in Victoria, where the Labor Party has held office for the past three consecutive terms. We condition LLMs with personas that are informed by census distributions. We then use news articles to provide the personas with political context, and prompt the models to simulate a complete ballot, including preferences, for each persona. We focus on the 88 seats in the Legislative Assembly. Although we provide an election forecast, our main contribution is identifying which factors systematically change various LLM outputs, and drawing attention to the aspects that should be disclosed when silicon samples are used for political forecasting.
