Traditional Methods Outperform Generative LLMs at Forecasting Credit Ratings