M2VN: Multi-Modal Learning Network for Volatility Forecasting

This repository contains the official implementation of M2VN, a lightweight yet effective architecture that combines price, volume, and news embeddings for equity–market volatility prediction.
The code accompanies our paper submitted to a ICAIF.

---

Repository Overview

Path / file	Purpose
data_provider/	Data loading and on-the-fly preprocessing
exp/	Experiment settings and logging utilities
layers/	Custom PyTorch layers used by M2VN
models/	Model definition and loss functions
runfile/	Shell scripts for training / inference (run_final.sh)
utils/	Miscellaneous helpers
Step 1–3 Aggregate Results.ipynb	Notebooks for reproducing paper tables
run.py	Entry point if you prefer python run.py over the shell script
LICENSE	License information (MIT)

---

Quick Start

1. Install packages

   pip install -r requirements.txt   # Provided in repo

2. Download the dataset

   The full, pre-processed dataset is available on Google Drive Link Also, raw version news artical data is here: Link

   After downloading, place the extracted folder inside e.g., (dataset/KO4.csv)

3. Train & evaluate

   bash ./runfile/run_final.sh

   The script trains M2VN with default hyper-parameters and writes:

   • Checkpoints → checkpoints/
   • Final metrics (CSV) → results_test/

---

Reproducing Paper Results

After training finishes, open the Jupyter notebooks in the repo root:

Notebook	What it does
Step 1 Aggregate Results.ipynb	Get Main model results
Step 2 Aggregate Results Vol(non).ipynb	W/O Volume
Step 3 Aggregate Results Vol-sent.ipynb	W/ News Sentiment

Running the cells in order will produce the tables reported in the paper.

---

Citation

If you find M2VN useful in your work, please consider citing:

@inproceedings{Anonymous,
  title     = {M2VN: Multi-Modal Learning Network for Volatility Forecasting},
  author    = {Anonymous},
  booktitle = {ICAIF},
  year      = {2025}
}

---

License

This project is licensed under the MIT License – see the LICENSE file for details.

---

Contact

Questions or suggestions? Open an issue or reach out to Anonymous.