Time Series Forecasting with LSTM and GRU Models

This repository contains Python code for time series forecasting using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks implemented in PyTorch. The models are trained and evaluated on a dataset named amazon.csv. (provided in this repository)

Features

• Model Architectures: Includes LSTM and GRU implementations for sequence modeling.
• Data Preprocessing: Time series data is prepared with scaling, sequence generation, and train-validation splitting.
• Model Training: Comprehensive training loop with loss computation and model validation.
• Performance Metrics: Training and validation loss are reported for each epoch to track model performance.

Requirements

• Python 3.8 or later
• PyTorch 2.0 or later
• NumPy
• Matplotlib
• scikit-learn

Install the required libraries using:

pip install torch numpy matplotlib scikit-learn

Hardware

• A CUDA-compatible GPU is recommended for efficient training.

---

Setup Instructions

1. Clone the Repository:

   git clone https://github.com/IAmFarrokhnejad/Stock-Forecasting-RNN

2. Install Dependencies:

   pip install torch torchvision matplotlib seaborn scikit-learn

3. Prepare Your Dataset:

• Ensure the amazon.csv file is placed in the root directory. The file should contain a single time series column for analysis.
• In case of changing the data directory, update the data variable path in the script to point to your dataset path.

Running the Code

Execute the script to train both the LSTM and GRU models:

python stock-forecasting.py

The script will:

1. Train and validate models sequentially for EfficientNet, MobileNetV2, and ShuffleNet.
2. Save the classification report and confusion matrix for each model in the specified directory.

---

Results

The models were trained for 40 epochs, and the following trends were observed:

1. LSTM:

• Training Loss decreased to 0.0001 by epoch 40.
• Validation Loss stabilized at 0.0016.

2. GRU:

• Training Loss decreased to 0.0001 by epoch 40.
• Validation Loss stabilized at 0.0019.

Both models demonstrate strong convergence, with GRU showing slightly slower validation loss improvement compared to LSTM.

---

License

This project is licensed under the MIT License.

---

Author

Morteza Farrokhnejad

---

Acknowledgments

• PyTorch for the deep learning framework.
• Matplotlib for visualization.