Forecasting is a cornerstone of data-driven decision-making, with recent advancements integrating contextual information and leveraging large language models (LLMs) to enhance and democratize predictions. However, are these innovations enough to ensure robust and reliable decision-making? In this talk, I will examine key limitations in current forecasting practices, drawing on insights from the Context is Key benchmark and the Regions of Reliability paper. Finally, I will highlight two overlooked aspects of complex systems that must be addressed to enable real-world decision support.