Term Selection for Stock Price Fluctuation Images by Large Mulitmodal Models

Abstract

In stock price fluctuation articles, stock terms are frequently used to describe stock price fluctuations. In order to automatically generate such articles, we evaluated a method for automatically selecting terms that appropriately represent the characteristics of stock price fluctuations by feeding a chart of stock price fluctuations over several days to large multimodal models. The method was found to be able to select, with high accuracy, terms that are manually assigned by journalists to articles on stock price fluctuations and terms that are close to those assigned by journalists. Comparative evaluation results also show that large multimodal models perform slightly worse but still mostly well compared with large language models such as GPT-4 (zero-shot and few-shot) and GPT-3.5 (fine-tuning).