Context information can be more important than reasoning for time series forecasting with a large language model