Beyond Naïve Prompting: Strategies for Improved Zero-shot Context-aided Forecasting with LLMs