Stock Price Prediction (LSTM vs GRU)

This project implements and compares two deep learning models — LSTM and GRU — to predict next-day Open and Close prices from historical OHLCV data. Models are built with PyTorch and trained on rolling 30-day sequences.

---

Models

• LSTM (Long Short-Term Memory): Captures long-term dependencies in time series.
• GRU (Gated Recurrent Unit): Lighter alternative with fewer parameters.

---

Features

• Data preprocessing with pandas, NumPy, scikit-learn
• Sequence windowing (past 30 days → next-day prediction)
• Training & inference with PyTorch
• Saved artifacts: lstm_model_weights.pth, scaler.save
• Performance comparison using MSE (per target)
• Visualizations of actual vs. predicted prices
• Ready to integrate with Streamlit (this repo) or Flask

---

Results

Model	MSE (Open)	MSE (Close)
LSTM	10.7422	16.8829
GRU	62.9326	50.9872

LSTM outperformed GRU in both Open and Close price predictions.

---

Project Structure

.
├── README.md
├── requirements.txt
├── app.py                  # Streamlit app
├── model.py                # LSTM/GRU definitions
├── lstm_model_weights.pth  # trained weights (PyTorch state_dict)
├── scaler.save             # fitted MinMaxScaler (joblib)
└── forecasting_notebook.ipynb  # training/experiments (optional)

---

Quickstart

1. Install dependencies

pip install -r requirements.txt

2. Run the Streamlit app

streamlit run app.py

3. Upload a CSV with columns:

Open, High, Low, Close, Volume

(Feature order must match the scaler used during training.)

---

Example Output

Actual Open: 111.20, Pred Open: 110.68 | Actual Close: 109.14, Pred Close: 111.37
Actual Open: 110.60, Pred Open: 110.17 | Actual Close: 110.73, Pred Close: 111.18
...

---

Deployment

• Load lstm_model_weights.pth and scaler.save
• Deploy with Streamlit (this repo) or Flask
• Optional: containerize with Docker for cloud deployment

---

Tech Stack

• Python
• PyTorch
• NumPy & Pandas
• scikit-learn
• Matplotlib & Seaborn

---

Live Demo

Try the model live on Streamlit: Stock Price Forecaster Demo

---

Author

Dante Santurian – University of Michigan