Large language models (LLM) or foundation models (FM) have been revolutionizing the scientific communities, including but not limited to math, physics, chemistry, biology, and material. However, the exploration of LLM in finance has been lagging behind. One bottleneck here is the lack of datasets for financial activities, and the other challenge is the unawareness of the incorporation of foundation models to explore versatile financial tasks. With this regard, we propose BudgetFM, a multi-modal large language model framework for financial activity predictions. BudgetFM tackles the first question by exploring a real-world dataset, Quebec Hotel: a hotel price and occupancy rate between Jan 2017 and Dec 2022, and the hotel is located in Quebec, Canada. Then BudgetFM handles the second question by utilizing the news data from heterogeneous resources of news, which can closely impact hotel budgeting. Empirical results verify the effectiveness of BudgetFM, with both the quantitative performance improvements (up to 95.04\%) and qualitative visualizations. Thus, we believe that BudgetFM brings a new direction in the large language model and finance communities. The code will be released following publication, and the data can be accessed upon request.