tobias-schnabel/mc-forecasting

Code for my Master's Dissertation in the MSc Statistical Science at the University of Oxford

Matrix Completion for Panel Data Forecasting

This repository contains the code and data for my master's thesis on using Matrix Completion methods for panel data forecasting, with a focus on electricity price forecasting. The dissertation, which includes my results, is available here.

Overview

This project explores the potential of Matrix Completion with Nuclear Norm Minimization (MC-NNM) for panel data forecasting. It implements and evaluates, for the first time, the autocorrelated errors extension to MC-NNM proposed by Athey et al. (2021), demonstrating its benefits in a forecasting context. See also my implementation of this estimator here.

Key Features

• New open-source Python implementation of MC-NNM with unprecedented flexibility in handling various covariate types
• Implementation of the autocorrelated errors extension (MC-NNM-TSR)
• Comprehensive comparison of MC-NNM against state-of-the-art electricity price forecasting methods
• Novel open-source forecast engine for standardized model comparison
• Rich dataset of European electricity prices for 12 countries retrieved from ENTSOE

Main Files

• data_loader.py: Functions for loading and preprocessing data
• data_utils.py: Utility functions for data manipulation
• estimator.py: Base class for estimator implementations
• forecast_engine.py: Main forecasting engine for model comparison
• models/: Folder containing individual model implementations
• training/: Scripts for training different models
• analysis/: Jupyter notebooks for results analysis and visualization

The forecast engine is designed to be easily customizable and extensible with new models. To add a new model, simply create a new class in the models/ folder that inherits from the Estimator base class and implements the fit and predict methods. It can handle any data, as long as it is passed in the correct format (see the docstrings).

Replication Instructions

To replicate my results:

1. Download the files in this repository at the tag Submission

wget https://github.com/tobias-schnabel/mc-forecasting/archive/refs/tags/Submission.zip
unzip Submission.zip
cd mc-forecasting-Submission

2. Install poetry if not already installed: pip install poetry
3. Install dependencies: poetry install
4. Run the two notebooks located in the analysis/ folder to reproduce the tables and figures
5. If you want to rerun the forecasts, which will take several days, you can re-run the forecasting experiments:

   for window in 56 84 112; do
    python training/train_elasticnet_${window}.py
    python training/train_lasso_${window}.py
    python training/train_lear_${window}.py
    python training/train_lear_uv_${window}.py
    python training/train_mc_${window}.py
    python training/train_mctsr_${window}.py
   done
   

Contributing

Contributions to improve the code, extend the analysis, or fix bugs are welcome. If you're interested in contributing, please follow these steps:

1. Fork the repository
2. Create a new branch for your feature (git checkout -b feature/AmazingFeature)
3. Make your changes
4. Commit your changes (git commit -m 'Add some AmazingFeature')
5. Push to the branch (git push origin feature/AmazingFeature)
6. Open a Pull Request

Code Style

Please adhere to the PEP 8 style guide for Python code. I recommend using an auto-formatter like Black or ruff to ensure consistency.