- Group Name: Karlis & Anas
- Group participants names: Bachiri Anas, Briedis Karlis Martins
- Project Title: The Impact of Network Topology on Banking Default Dynamics
- Programming language: Python 3
Systemic risk is of crucial importance when dealing with complex systems. In particular, financial networks are an important illustration of such systems. In this case, financial institutions interacts with each others respecting the graph structure of the underlying network.
Recent studies have revealed the relevance of financial network structure in estimating systemic risk. It has also showed that the neglect of the graph topology is seriously underestimating the risk of default of the whole network. In this project, we address the issue of crisis propagation in a banking network. We first construct a banking system with some parameters, distribute the assets on the banks and finally we simulate shocks in this financial network. This work investigate the general behaviour of crisis propagation in a financial network, and therefore provides some insights about reducing the systemic risk. For this, we will scan the parameters of the used model and describe the behaviour of the resilience of financial networks.
The model implemented in our project is based on a model provided by Nier and collaborators. The latter uses random networks and some aggregated parameters of assets to create a financial network. Then it performs shocks within the same network and propagate the initial shock. Graph with advanced topological properties like average clustering coefficient are generated using simulated annealing evolutionary algorithm and simulation is ran to observe how these properties affects default dynamics.
The main goal of this project is to investigate how does the topology of complex networks impact the default dynamics of the banking network.
More specific questions include:
- How does the network density impact the number of defaults of financial network?
- How does the clustering of network impact the number of defaults of financial network?
- etc.
The project in mainly based on these two papers:
- Nier, Erlend & Yang, Jing & Yorulmazer, Tanju & Alentorn, Amadeo. (2008). Network Models and Financial Stability. Journal of Economic Dynamics and Control. 31. 2033-2060. 10.1016/j.jedc.2007.01.014.
- V. Kashirin, Victor. (2014). Evolutionary Simulation of Complex Networks Structures with Specific Topological Properties. Procedia Computer Science. 29. 2401-2411. 10.1016/j.procs.2014.05.224.
Random graph generation of simulated annealing. Crisis propagation in random graph.
Prerequisite for all reproducibility tests is
python >=3.6
python -m pip install numpy matplotlib python-igraph
For installation of igraph, please refer to its python-igraph Manual. For Windows users a convinient way is to use unofficial windows binaries by downloading suitable
.whlfile and running e.g.pip install python_igraph‑0.7.1.post6‑cp36‑cp36m‑win_amd64.whl
git clone https://github.com/kmbriedis/msss_project.git
cd msss_project/code
To run light tests, make sure all dependencies are installed and you are in cloned directory.
To reproduce results of Nier et al. and observe how default dynamics of banking network are affected by variation of net worth, interbank assets, or Erdös-Rényi probability, run (running time 1-2 minutes):
python test_light.py --reproduce-sim
You should see graphs like these, which show how the default dynamics (number of defaulted banks) are affected by variantion of different attributes
| Light test | Full test | Nier et al. | |
|---|---|---|---|
| Percentage net worth variation |  |
 |
 |
| Percentage interbank assets variation |  |
 |
 |
| Erdös-Rényi probability variation |  |
 |
 |
There are pregenerated graphs in other/pregenerated_graphs directory. Unzip clustering_light.zip and communities.zip in your chosen $UNZIP_DIR
Run (by replacing $UNZIP_DIR with extraction directory):
python test_light.py --sim-communities $UNZIP_DIR/communities
python test_light.py --sim-clustering $UNZIP_DIR/clustering_light
You should see graphs like these, which show how the default dynamics (number of defaulted banks) are affected by different clustering coefficients and number of communities
| Light test | Full test | |
|---|---|---|
| Clustering coefficient variation |  |
 |
| Number of communities variation |  |
Same as light |
Make sure all dependencies are installed and you are in the cloned directory.
Refer to Light test, but simulations must be run with command python test_light.py --reproduce-sim, where sample points and samplea are increased (10x increase in number of graphs and running time).
Refer to Light test, but running the command with pregenerated graphs from other/pregenerated_graphs/clustering.zip
To pregenerate graphs for further simulation, run these commands
Warning Generation is compute-intensive and can take up to 6 hours when running on ETH's Euler cluster's 48 cores (12 days of CPU time) when using --clustering flag.
python generate_multicore.py --communities
python generate_multicore.py --clustering-light
python generate_multicore.py --clustering
To run the generation on ETH Zurich Euler cluster, use SSH to access the cluster, connectiong to <nethz_username>@euler.ethz.ch and run following commands:
mkdir -p $HOME/python/lib64/python3.6/site-packages
export PYTHONPATH=$HOME/python/lib64/python3.6/site-packages:$PYTHONPATH
module load python/3.6.0
python -m pip install --install-option="--prefix=$HOME/python" python-igraph numpy
git clone https://github.com/kmbriedis/msss_project.git
cd msss_project/code
bsub -W 24:00 -n 48 python generate_multicore.py --clustering # or any other command