Comparison of an affine term structure model with Fed chair speeches in large language models

We compare the performance of financial sentiment analysis on Fed chair speeches between a domain-specific model (finBERT) and a general-purpose model (flan-T5) with few-shot learning. Specifically, we implement an out-of-sample yield forecast of an affine term structure model using two different sets of sentiment factor values for Fed chair speeches and compare the root mean squared error (RMSE) and the mean absolute deviation (MAD) of the yield forecasts between two sentiment analysis models. The performance of the general-purpose model with few-shot learning seems comparable to the domain-specific model. Considering the computational costs of pre-training and fine-tuning a domain-specific model, it seems cost-efficient to use general-purpose models with few-shot learning for the sentiment analysis of the Fed chair speeches.