GenChaR: A Dataset for Stock Chart Captioning

In this work, we introduce a new dataset GenChaR for an image captioning task around stock charts. The task aims to read market sentiment directly from depicted charts and generate descriptions, hopefully to provide comprehensible and useful insights for stock trading. Impressed by the success of large language models (LLMs), the study decides to pioneer itself by exploring the capabilities of large vision-language models (LVLMs) on the proposed task. This paper outlines the objectives of the stock captioning task, the dataset we built, and automatic evaluation with some representative general-purpose LVLMs.