The first thing I'd do if working with an LLM on tabular data is to ask what the best tool would be to work with that data and build up a proper harness to work with the data sensibly. Rawdogging LLM isn't the tool for forecasting like this, as they found.