StockGenChaR: A Study on the Evaluation of Large Vision-Language Models on Stock Chart Captioning

TL;DR

This paper introduces StockGenChaR, a new dataset for evaluating large vision-language models on stock chart captioning, aiming to assist investors by generating market sentiment descriptions from stock chart images.

Contribution

The study presents a novel dataset and evaluation framework for large vision-language models applied to stock chart captioning, bridging finance and AI.

Findings

Large vision-language models can generate market sentiment descriptions.

StockGenChaR dataset enables benchmarking of captioning performance.

Potential to assist non-expert investors with automated chart interpretation.

Abstract

Technical analysis in finance, which aims at forecasting price movements in the future by analyzing past market data, relies on the insights that can be gained from the interpretation of stock charts; therefore, non-expert investors could greatly benefit from AI tools that can assist with the captioning of such charts. In our work, we introduce a new dataset StockGenChaR to evaluate large vision-language models in image captioning with stock charts. The purpose of the proposed task is to generate informative descriptions of the depicted charts and help to read the sentiment of the market regarding specific stocks, thus providing useful information for investors