Predicting Systemic Risk in Financial Systems Using Deep Graph Learning

Repo for the paper: “Predicting Systemic Risk in Financial Systems Using Deep Graph Learning”

Code for systemic risk classification and percentile prediction using Graph Neural Networks and Class to Regression (C2R).

Installation

1. Clone this repository (Python 3.9)

   git clone https://github.com/vibalcam/gnn-systemic-risk.git

2. Install Pytorch and DGL library.

3. Install Python dependencies

   pip install -r requirements.txt

Code Organization

Results

Inside the notebooks folder:

• results.ipynb contains the code to summarize the results
• graphs folder contains visualizations for the networks
• results_cm folder contains the confusion matrices for the classification models
• results_conf folder contains visualizations for the confidence intervals

Notebooks

Each network has a folder with the different scenarios, the best model for each type, and a notebook (models_training.ipynb) with the models' training.

Data generation

The code that generates the networks from the aggregated data can be found in each of the network's folders (inside the notebooks folder) by the name generate_data.R. Each folder also contains a generate_data.RData file with the saved workspace.

The networks themselves can be found inside the data folder. Each network is divided into two files. The file network.csv contains the adjacency matrix of the network, and nodes.csv contains the nodes features.

Models

The code for the models and training can be found in the models folder. The models.py file contains the model definitions, train.py contains the code to train the classification models, and train_reg.py contains the code to train the percentile regression models.

Reference

If you find it helpful, please cite our paper:

@article{balmaseda_predicting_2023,
	title = {Predicting systemic risk in financial systems using {Deep} {Graph} {Learning}},
	volume = {19},
	issn = {2667-3053},
	url = {https://www.sciencedirect.com/science/article/pii/S2667305323000650},
	doi = {https://doi.org/10.1016/j.iswa.2023.200240},
	journal = {Intelligent Systems with Applications},
	author = {Balmaseda, Vicente and Coronado, María and de Cadenas-Santiago, Gonzalo},
	year = {2023},
	keywords = {Financial networks modeling, Graph neural networks (GNN), Label regression, Model selection, Network simulation, Neural networks},
	pages = {200240},
}