Transformer-Based Multi-Modal Stock Market Prediction

Research project exploring short-horizon stock forecasting by combining time-series signals with textual and macroeconomic context.

The repository compares multiple modeling approaches across notebooks and focuses on whether adding richer contextual inputs, such as sentiment and market indicators, can improve prediction quality over purely price-based baselines.

Project Focus

• Forecast short-term stock movement and price behavior
• Compare classical and deep learning approaches
• Incorporate textual information and market context into the modeling pipeline
• Evaluate how different modeling choices affect forecast quality and stability

Signals Used

The experiments in this repository draw from several types of input:

• historical stock prices
• ticker-level sequences
• VIX and market context
• textual summaries and sentiment-style signals

Modeling Tracks

• SARIMA.ipynb: classical statistical baseline
• LSTM.ipynb: recurrent neural network baseline
• T5.ipynb: sequence-to-sequence forecasting experiments
• TFT.ipynb, TFT_1.ipynb, TFT_Phase1.ipynb, TFT_Phase2.ipynb, TFT_Text.ipynb: Temporal Fusion Transformer experiments, including multimodal and sentiment-aware variants

Repository Contents

• CSCI_566___Final_Project_Report.pdf: project report
• 65_G.O.A.T.pdf: additional report artifact
• processed_vix.csv: processed macro/market signal input
• all_tickers (1).txt: ticker universe reference
• notebook-based experimentation and evaluation files

How To Explore

1. Start with CSCI_566___Final_Project_Report.pdf for the research framing.
2. Review SARIMA.ipynb and LSTM.ipynb for baseline approaches.
3. Move to the TFT* notebooks to see the richer multimodal forecasting work.
4. Inspect T5.ipynb for alternative sequence-model experimentation.

Tech Stack

• Python
• Jupyter Notebook
• PyTorch Forecasting
• Transformers
• statsmodels
• yfinance
• pandas and NumPy

Why This Repo Matters

This project demonstrates applied ML research thinking: comparing baselines, layering in richer signals, and using multiple model families to study a difficult forecasting problem rather than relying on a single technique.