Comparing Vision-Instruct LLMs, Vision-Based Deep Learning, and Numeric Models for Stock Movement Prediction

This research conducts a comparative study of several stock movement prediction approaches, evaluating large language models (LLMs) and vision-based deep learning models with stock image as input, as well as models that utilize numerical data. Specifically, the study investigates a prompt-based LLM framework that processes candlestick charts, comparing its performance with image-based models such as MobileNetV2, Vision Transformer, and Convolutional Neural Network (CNN), as well as models with numerical inputs including Support Vector Machine (SVM), Random Forest, LSTM, and CNN-LSTM. Although LLMs have demonstrated promising results in stock prediction, directly applying them to stock images poses challenges compared to numerical approaches. To address this, this study further improves LLM performance with post-hoc calibration, reducing prediction biases. Experimental results demonstrate that post-hoc calibrated LLMs with visual input achieve competitive performance compared to other models, highlighting their potential as a viable alternative to traditional stock prediction methods while simplifying the prediction process.

@article{Chen2025,
  title     = {Comparing Vision-Instruct LLMs, Vision-Based Deep Learning, and Numeric Models for Stock Movement Prediction},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {4},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Qizhao Chen},
  doi       = {10.14569/IJACSA.2025.0160402},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160402}
}