Time Series Stock Price Prediction using LSTM Model

Table of Contents

• Project Overview
• Dataset Overview
• Data Preprocessing
• Model Architecture
• Model Training and Evaluation
• Results and Discussion
• Conclusion
• License

Project Overview

This project focuses on predicting stock prices of Apple (AAPL) and AMD (AMD) using LSTM (Long Short-Term Memory) neural networks. The goal is to forecast future stock prices based on historical stock data. The project follows a step-by-step approach, from data exploration and preprocessing to building and evaluating multiple LSTM models. Modifications like adding dropout, regularization, and fine-tuning hyperparameters are explored to improve the model performance.

Dataset Overview

The datasets for this project contain historical stock data for Apple (AAPL) and Advanced Micro Devices (AMD):

• Apple (AAPL): The dataset includes daily stock prices from 1980 to 2020.
• AMD (AMD): The dataset includes daily stock prices from 1980 to 2020.

Each dataset consists of columns such as Date, Open, High, Low, Close, Adjusted Close, and Volume.

Data Preprocessing

Data preprocessing steps include:

1. Rescaling: Data is normalized using Min-Max scaling to bring all values into the range [0, 1], which helps improve model convergence.
2. Windowing: The time series data is split using a window size of 5, meaning the model will use data from 5 days to predict the price on the 6th day (next day).
3. Splitting the Dataset: The dataset is split into:
   • 80% training data
   • 10% validation data
   • 10% test data

The datasets for both AAPL and AMD are prepared separately for training and evaluation.

Model Architecture

Two models are implemented for stock price prediction:

1. Baseline Model: The first model is a simple LSTM architecture with 50 units and one output layer. It uses ReLU activation for the LSTM layer and Adam optimizer.
2. Modified Model: The second model enhances the baseline by adding a second LSTM layer, dropout regularization (with a dropout rate of 0.2), L2 regularization, and tuning hyperparameters such as learning rate and epochs.

Model Parameters:

• LSTM Layers: The first LSTM layer has 50 units, and the second LSTM layer has 30 units.
• Dropout: Dropout layers are added to reduce overfitting.
• L2 Regularization: Used on both LSTM and Dense layers to prevent overfitting.

Model Training and Evaluation

Training is carried out using Adam optimizer, and the models are evaluated on the test set using metrics like:

• Root Mean Squared Error (RMSE)
• Mean Absolute Error (MAE)
• Mean Absolute Percentage Error (MAPE)

Evaluation Results:

• Baseline Model: Trained for 50 epochs, but there was some overfitting, especially with volatile stock price data.
• Modified Model: Achieved better performance with improved generalization and stability due to the additional LSTM layer, dropout, and regularization techniques.

Results and Discussion

Model Performance

• Baseline Model: The baseline model for AAPL and AMD had a relatively good performance in following general trends but faced issues like overfitting, especially during periods of high volatility.
• Modified Model: The second model with modifications like dropout and additional LSTM layers performed better. The validation loss became more stable, and the model managed to capture more complex patterns in stock data.

Key Insights

1. Model Stability: The modified model improved stability, showing less fluctuation in validation loss compared to the baseline model.
2. Overfitting: Both models showed overfitting to some extent, especially during volatile periods, which indicates the need for more regularization.
3. Stock Price Volatility: Despite the improvements, the models faced difficulty in predicting extreme price changes, a common issue in financial markets.

Limitations

• Overfitting: While the modified model improved performance, overfitting remained a challenge. Regularization techniques like L2 helped, but further improvements could be made with more data or advanced techniques.
• Stock Market Volatility: The models struggled with periods of high volatility, where stock prices can change unpredictably, making predictions more difficult.

Conclusion

This project demonstrates the potential of LSTM networks in predicting stock prices, especially for data with temporal patterns. While both the baseline model and the modified model performed well in capturing general trends, the modified model proved to be more effective in generalizing to unseen data. Future work could involve exploring other architectures like GRU, increasing data volume, and applying more sophisticated techniques to handle volatility in the stock market.

License

This project is licensed under the MIT License - see the LICENSE file for details.