🧠 NVDA LSTM Predictor

A deep learning model that predicts whether NVIDIA's stock (NVDA) will go UP or DOWN next week based on historical data and technical indicators - using an LSTM (Long Short-Term Memory) neural network trained on normalized weekly price patterns.

NVDA-LSTM-PREDICTOR/
│
├── assets/
│   ├── split_not_taken_into_account.png
│   ├── split_taken_into_account.png
│   └── training_vs_validation_accuracy.png
│
├── NVDA_LSTM_Predictor.ipynb    # Jupyter Notebook with all code
├── requirements.txt             # Python dependencies
├── .env                         # Alpaca API keys (not committed)
├── README.md                    # Project documentation
├── .gitignore
└── venv/                        # Virtual environment (excluded from Git)

We trained a neural network to predict whether the next week's closing price of NVDA stock will be higher than this week's, using:

• Weekly historical stock data from Alpaca API
• Technical indicators: MACD, RSI, SMAs
• Preprocessing: stock split adjustment, normalization, sequence generation
• A custom LSTM model trained with binary_crossentropy loss

Because the stock market is noisy and volatile daily, we shift our perspective to weekly resolution, making it more reliable for trend-based strategies.

All indicators are calculated using the ta library.

In June 2024, NVDA had a 10:1 stock split. Without adjusting historical data, the model would learn garbage signals.

Visual comparison:

Adjusted for Split ✅	Not Adjusted ❌

We used a custom adjust_for_stock_split() function to scale historical prices and volumes appropriately.

LSTMs (Long Short-Term Memory networks) are specialized RNNs designed for sequence prediction. They remember patterns across time - perfect for stock price time series.

• Input: Past 10 weeks of technical indicators
• Output: A binary prediction for next week: 1 (UP) or 0 (DOWN)
• Architecture:
  • 2 stacked LSTM layers (64 units each)
  • Dropout + L2 regularization to prevent overfitting
  • Dense sigmoid output layer

Train accuracy: 65.00%
Test accuracy: 64.00%
Validation Gap: 1.00%

The model generalizes well and performs significantly better than random guessing (50%).

Prediction for tomorrow:
Probability of price increase: 61.39%
Predicted direction: UP

git clone https://github.com/MisbahAN/NVDA-LSTM-PREDICTOR.git
cd NVDA-LSTM-PREDICTOR

pip install -r requirements.txt

APCA_API_KEY_ID=your_key_here
APCA_API_SECRET_KEY=your_secret_here

Get your keys for free from Alpaca Markets.

jupyter notebook NVDA_LSTM_Predictor.ipynb

• Uses real-time stock data from Alpaca
• Automatically adjusts for stock splits
• Easy-to-read Plotly visualizations
• Binary direction prediction via LSTM
• Achieves ~64% test accuracy with minimal overfitting
• Modular structure for reuse on any stock

Misbah Ahmed Nauman
🌐 MisbahAN.com