A financial multivariate time series forecasting framework incorporating large language models and its experimental evaluation

By analyzing mainstream time series forecasting methods based on large language model （LLM）, a unified model framework is proposed and empirically evaluated on exchange rate and stock index data. Experimental results indicate that LLM exhibit certain performance advantages in financial time series forecasting, but also face notable limitations. Especially, relying solely on simple textual inputs or prompts may not lead to performance improvements.