macroRL: LSTM-Based Macroeconomic Forecasting with PyTorch

This repository contains a toy project for forecasting macroeconomic variables (e.g., CPI-based inflation) using an LSTM in PyTorch. It includes TensorBoard logging for visualization, and fetches data from FRED via pandas_datareader.

Table of Contents

1. Prerequisites
2. Installation Using Miniconda (Recommended on Windows)
3. Usage
4. Running TensorBoard
5. Common Issues and Troubleshooting
6. License

---

Prerequisites

• Miniconda (or Anaconda) on Windows, macOS, or Linux.
• Basic familiarity with Python, PyTorch, and Git.

If you prefer a plain Python installation without Conda, you must manually manage your environment and ensure all dependencies are installed (PyTorch, pandas, numpy, etc.). Using Conda simplifies environment management and avoids version conflicts.

---

Installation Using Miniconda (Recommended on Windows)

1. Install Miniconda

• Download Miniconda for your platform from:
  https://docs.conda.io/en/latest/miniconda.html
• For Windows, choose “Windows 64-bit” (the default on most modern systems).
• Double-click to install, following the prompts. You can keep the default settings.

2. Launch the Miniconda Prompt

• On Windows, open your Start Menu and look for "Miniconda Prompt" (or "Anaconda Prompt").
• This special prompt ensures conda is recognized.

3. Clone or Download this Repository

• If you have Git installed, open the Miniconda Prompt and run:

  git clone https://github.com/YOUR_USERNAME/YOUR_REPO.git
  cd YOUR_REPO

• Or download the repo as a ZIP from GitHub and extract it, then cd into the unzipped folder in the Miniconda Prompt:

  cd path\to\YOUR_REPO

4. Create a Conda Environment

conda create -n macroRL python=3.10
conda activate macroRL

• This creates an environment named macroRL with Python 3.10.
• Adjust the Python version if you wish (e.g. 3.9, 3.11).

5. Install Dependencies

• CPU-Only PyTorch (simplest if you don’t need GPU):

  conda install pytorch torchvision torchaudio cpuonly -c pytorch

• NVIDIA GPU + CUDA (for faster training):

  conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

  This automatically installs CUDA 11.8 support with PyTorch.

• Other libraries:

  conda install pandas numpy pandas-datareader
  pip install tensorboard

  (Optionally add conda install matplotlib if you want local plots.)

6. Verify Your Environment

• Check Python:

  python --version

• Check PyTorch version:

  python -c "import torch; print(torch.__version__)"

• Check TensorBoard import:

  python -c "import tensorboard"

If these commands run without errors, you’re set.

---

Usage

1. Activate the Environment
   In a new terminal (Miniconda Prompt):

   conda activate macroRL

   Then navigate to the project folder:

   cd path\to\YOUR_REPO

2. Run the Training Script

   python main.py

   • This script will:
     1. Fetch CPI data from FRED (via pandas_datareader).
     2. Compute year-over-year inflation.
     3. Train an LSTM model in PyTorch.
     4. Log metrics (loss, RMSE, etc.) and plots to TensorBoard in a runs/ directory.

3. (Optional) Adjust Hyperparameters

   • Open main.py and change parameters like EPOCHS, LOOKBACK, or the LSTM hidden_size.
   • Rerun python main.py to see how it affects training.

---

Running TensorBoard

To view logs in real-time or after training completes:

conda activate macroRL
cd path\to\YOUR_REPO
tensorboard --logdir=runs --port=6006

If tensorboard isn’t recognized, try:

python -m tensorboard --logdir=runs --port=6006

Open http://localhost:6006 in your browser.
You’ll see:

• Scalars: Training/validation losses, RMSE, MAE across epochs
• Figures: Plots of actual vs. predicted inflation (if your script logs them)
• Graphs: High-level model graph (optional)

---

Common Issues and Troubleshooting

1. conda: command not found or “Term 'conda' is not recognized”

   • On Windows, ensure you’re in the Miniconda Prompt (Start Menu → “Miniconda3 Prompt”) or have run conda init powershell so that conda is recognized in your shell.
   • Restart your terminal/PowerShell if changes have been made to PATH.

2. No module named tensorboard or “tensorboard not recognized”**

   • Install TensorBoard again with:

     conda activate macroRL
     pip install tensorboard

   • Or run with python -m tensorboard --logdir=runs.
   • Confirm you’re in the correct conda environment.

3. FRED Data Access Problems

   • If pandas_datareader cannot fetch data due to rate limits, you may need an API key. See FRED API docs.
   • Or manually download a CSV from FRED and load it in main.py instead.

4. GPU Not Detected

   • Check if your GPU is recognized by PyTorch:

     python -c "import torch; print(torch.cuda.is_available())"

   • If False, ensure you installed the GPU version (pytorch-cuda=11.8) and have compatible NVIDIA drivers.
   • Make sure your script sets DEVICE = "cuda" (if torch.cuda.is_available() returns True).

---

License

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