Stock Trend Predictor

A LSTM-based machine learning project for predicting stock price trends using historical data from Yahoo Finance. This repository includes training scripts, Jupyter notebooks for model development, and a Streamlit web app for interactive visualization and prediction.

Features

• Data Collection: Downloads historical stock data using yfinance for any ticker symbol.
• Data Preprocessing: Handles MultiIndex columns from yfinance, computes moving averages, percentage changes, and scales data using PyTorch-only MinMaxScaler.
• Model Training: Builds and trains an LSTM neural network using PyTorch to predict closing prices based on past 100 days of data.
• Evaluation: Computes metrics like MAE, RMSE, MAPE, and R2; plots actual vs predicted prices.
• Web App: Streamlit interface for uploading models, selecting stocks, and visualizing predictions.
• Model Persistence: Saves trained models as .pth files for reuse.
• Pure PyTorch: No scikit-learn dependencies - all preprocessing and scaling done with PyTorch.

Project Structure

Stock-Trend-Predictor/
├── main.py                 # Streamlit web app for prediction and visualization
├── src/
│   └── stock_predictor/    # Main package
│       ├── __init__.py     # Package initialization
│       ├── model.py        # PyTorch LSTM model architecture
│       └── scaler.py       # PyTorch-based MinMaxScaler implementation
├── scripts/
│   └── train.py            # Training script for PyTorch model
├── notebooks/
│   └── stock.ipynb         # Jupyter notebook for data exploration, training, and evaluation
├── model.pth               # Pre-trained PyTorch model (generated after training, not included in repo)
├── requirements.txt        # Python dependencies
├── .gitignore              # Git ignore file
└── LICENSE                 # License file

Installation

1. Clone the repository:

   git clone https://github.com/bedigambar/Stock-Trend-Predictor-Model
   cd Stock-Trend-Predictor-Model

2. Create a virtual environment (recommended):

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

Training a Model

Train a new PyTorch model using the training script:

python scripts/train.py

This will:

• Download 20 years of GOOG stock data
• Preprocess and scale the data using PyTorch MinMaxScaler
• Train an LSTM model for 20 epochs
• Save the model as model.pth

To customize training:

• Edit stock_symbol in scripts/train.py to change the stock
• Adjust epochs to change training duration
• Modify seq_len to change sequence length (default: 100)

Running the Streamlit App

The Streamlit app allows you to load stock data, upload or auto-detect a trained model, and run predictions.

streamlit run main.py

• Enter a stock symbol (e.g., "GOOG").
• Adjust years of data and prediction settings.
• Upload a .pth model file or let it auto-detect model.pth.
• View plots and metrics.

Running the Jupyter Notebook

Use the notebook for development, training, and evaluation.

1. Open notebooks/stock.ipynb in Jupyter Lab or VS Code.
2. Run cells in order:
   • Import libraries and download data.
   • Preprocess data (normalize columns, compute features).
   • Prepare sequences for LSTM (100-day windows).
   • Train the PyTorch model.
   • Evaluate and save the model.
   • Load the saved model and re-evaluate.

Key cells:

• Data download: Uses yf.download() with MultiIndex normalization.
• Model architecture: LSTM(128) -> LSTM(64) -> Linear(25) -> Linear(1).
• Training: PyTorch training loop with Adam optimizer.
• Evaluation: Computes MAE, RMSE, MAPE, R2 and plots results.

Model Details

• Architecture: PyTorch LSTM model with two LSTM layers (128 and 64 hidden units), followed by linear layers.
• Input: Sequences of 100 days of scaled closing prices (shape: (batch_size, 100, 1)).
• Output: Predicted closing price for the next day.
• Training: Adam optimizer (lr=0.001), MSE loss, batch size 32.
• Preprocessing: PyTorch-based MinMaxScaler (no scikit-learn dependency).
• Evaluation Metrics:
  • Mean Absolute Error (MAE)
  • Root Mean Square Error (RMSE)
  • Mean Absolute Percentage Error (MAPE)
  • R-squared (R2)

Dependencies

See requirements.txt for the full list. Key packages:

• torch: PyTorch for model building and training.
• streamlit: Web app framework.
• yfinance: Data download.
• pandas, numpy: Data handling.
• plotly: Interactive plotting.
• matplotlib: Additional plotting support.

Notes

• This project uses pure PyTorch - no scikit-learn or TensorFlow/Keras dependencies.
• The src/stock_predictor/scaler.py module provides a drop-in replacement for sklearn's MinMaxScaler.
• Models are saved as .pth files using PyTorch's state_dict format.
• The project follows a clean package structure with organized directories for better maintainability.

License & Attribution

Stock Trend Predictor is released under the MIT License — © 2026 bedigambar.

You're welcome to use, modify and build on this code, but you must keep the copyright notice and license intact (i.e. give credit). If you ship something based on Stock Trend Predictor, a shout-out and a link back to this repo is hugely appreciated. 🙏

Developed by Digambar.