anubhav0907/Nifty50-LSTM

Nifty 50 Multi-Model Stock Price Prediction

A deep learning project for Nifty 50 forecasting using multiple neural architectures (LSTM, GRU, Transformer, and CNN+LSTM), with a Streamlit dashboard for model comparison, forecasting, and interactive technical analysis.

Project Structure

Stock Prediction/
└── Data  collection/
    ├── data_fetch.ipynb        # Data download & feature engineering
    ├── model_train.ipynb       # Multi-model training & evaluation
    ├── predict.ipynb           # Multi-model inference + best-model forecasting
    ├── app.py                  # Streamlit dashboard
    ├── data/
    │   ├── nifty50_raw.csv     # Raw OHLCV data (2018–2024)
    │   ├── processed_data.pkl  # Scaled sequences + scalers (dict format)
    │   └── forecast_30days.csv # 30-day forecast output
    └── models/
        ├── lstm_model.h5             # LSTM checkpoint
        ├── gru_model.h5              # GRU checkpoint
        ├── transformer_model.keras   # Transformer checkpoint
        └── cnn_lstm_model.keras      # CNN+LSTM checkpoint

Features

• —11 input features: Close, Open, High, Low, Volume, MA20, MA50, RSI, Bollinger Bands (upper/lower), Returns
• —4 trained architectures: LSTM, GRU, Transformer, CNN+LSTM
• —Chronological validation with shuffle=False for time-series-safe training
• —Automatic model leaderboard (RMSE, MAE, MAPE) in prediction notebook and dashboard
• —Best-model-driven forecasting for 7 to 60 business days
• —Apple-inspired Streamlit UI (clean light theme, SF-style typography, premium card layout)
• —Model benchmark sections with rolling MAPE, metric bars, pairwise model-distance heatmap, and baseline comparisons
• —Streamlit dashboard with model selection, comparison overlays, forecast download, and technical indicators

Model Architecture Summary

• —LSTM: 2 stacked LSTM layers (48 -> 24), Dropout, Dense head
• —GRU: 2 stacked GRU layers (48 -> 24), Dropout, Dense head
• —Transformer: MultiHeadAttention + feed-forward block + dense regression head
• —CNN+LSTM: stacked Conv1D blocks + deep LSTM stack + dense head

Common setup:

• —Loss: Huber
• —Optimizers: Adam with gradient clipping (clipnorm=1.0)
• —Window size: 60 days
• —Train/Test split: chronological (time-series-safe)

Performance

Performance is model-dependent and can change with retraining. Use:

• —predict.ipynb for a sorted model leaderboard (RMSE, MAE, MAPE)
• —Streamlit dashboard tab Model Performance for active model diagnostics and full-model leaderboard

Setup & Usage

1. Install dependencies

pip install tensorflow scikit-learn pandas numpy matplotlib yfinance streamlit plotly

2. Fetch data & build features

Run all cells in data_fetch.ipynb — saves data/processed_data.pkl.

3. Train models

Run all cells in model_train.ipynb — saves model checkpoints in models/:

• —lstm_model.h5
• —gru_model.h5
• —transformer_model.keras
• —cnn_lstm_model.keras

4. Run prediction workflow

Run predict.ipynb to:

• —Load all available trained models
• —Evaluate each model on test data
• —Auto-select the best model (lowest MAPE)
• —Generate and save data/forecast_30days.csv with model name metadata

5. Run the Streamlit dashboard

python -m streamlit run app.py

The dashboard now:

• —Loads all available models from Data collection/models
• —Lets you choose the active model in the sidebar
• —Supports optional prediction overlay for all models
• —Uses selected model for metrics and forecast download
• —Includes a dedicated Model Benchmarks tab for model-vs-model analysis

6. Deploy on Hugging Face Spaces (Docker)

1. 1.Create a new Hugging Face Space and choose Docker as the SDK.
2. 2.Point the Space to this GitHub repository.
3. 3.Hugging Face builds from Dockerfile and runs root app.py, which forwards to Data collection/app.py.
4. 4.After each push to main, the Space rebuilds and redeploys.

Tech Stack

• —Python 3.12
• —TensorFlow / Keras
• —scikit-learn
• —pandas, NumPy
• —Streamlit
• —Plotly
• —yfinance