Describing financial crisis propagation through epidemic modelling on multiplex networks

Malvina Bozhidarova, Frank Ball, Yves van Gennip, Reuben D. O’Dea, Gilles StupflerProceedings of the Royal Society A Mathematical Physical and Engineering Sciences 2024 · DOI

TL;DR

This paper introduces a novel framework for modeling financial crises as epidemics spreading through multiplex networks. By combining stock market data, geographical information, and economic sectors, the model effectively predicts crisis propagation, outperforming simpler models and highlighting the importance of diverse network structures.

What Problem Does This Paper Address?

Financial crises have become increasingly frequent and impactful, highlighting the critical need to understand how they spread through interconnected global markets. Traditional models often struggle to capture the complex, multi-faceted nature of these contagions. This paper addresses this gap by proposing a more sophisticated approach that accounts for various types of relationships between companies, moving beyond simple one-to-one connections to model how a crisis in one entity can cascade through different layers of influence.

Methods

The core of this research lies in a novel framework that integrates three key components:

  • Multiplex Network Construction: The authors build complex networks where each company is a node. These networks are "multiplex," meaning they have multiple layers, each representing a different type of connection or influence. These layers include:
  • Tail Dependence Network: This layer captures the strength of extreme co-movements in stock prices between companies, using Extreme Value Theory (EVT) to quantify the likelihood of simultaneous large losses.
  • Continents Layer: Companies within the same geographical continent are connected, reflecting regional economic ties.
  • Sectors Layer: Companies in the same economic sector are connected, acknowledging industry-specific vulnerabilities.
  • Global Layer: This layer represents "pure contagion," where any company can potentially influence any other, regardless of direct links, reflecting global market sentiment or systemic shocks.
  • Epidemic Modeling (SIR Model): A discrete-time Susceptible-Infected-Recovered (SIR) model, commonly used in epidemiology, is applied to these multiplex networks. Companies can transition from a "susceptible" state to "infected" (experiencing a financial shock) and then to "recovered." The model simulates the spread of "infection" based on the network structure and estimated transmission probabilities.
  • Parameter Estimation: Transmission probabilities between companies are estimated using a maximum likelihood approach based on historical stock price data.

Datasets

The study utilizes historical daily stock price data for 398 companies spanning from January 17, 2002, to July 18, 2022. This data was collected from Yahoo Finance. The companies were grouped by geographical location and primary economic sector, as defined by the Bureau van Dijk dataset. The analysis specifically focuses on two distinct periods to model the 2008 and 2020 financial crises.

Key Insights and Ideas

  • Multiplexity Matters: The most significant insight is that financial contagion is not a monolithic phenomenon. Different types of connections (e.g., direct financial links, geographical proximity, sectoral commonalities, and global sentiment) play distinct roles in how crises spread.
  • Tail Dependence as a Key Indicator: The "tail dependence" layer, derived from EVT, proves crucial. It captures the tendency for extreme events in one company's stock to be associated with extreme events in another, providing a more nuanced view of interconnectedness than simple correlation.
  • Beyond Homogeneous Mixing: The research strongly demonstrates that assuming all companies are equally likely to interact (homogeneous mixing) is a poor approximation for financial markets. The layered network structure significantly improves prediction accuracy.
  • Layered Importance: While the full multiplex model performs best, the study reveals that even simpler combinations of layers, particularly those including the tail dependence (PMFG) layer alongside the global layer, offer substantial predictive improvements over the basic global layer alone.

Key Results

  • Improved Prediction Accuracy: The proposed multiplex network SIR model consistently outperforms a homogeneous mixing population model in predicting the number of infected companies, affected sectors, and geographical locations.
  • Full Model Superiority: The model incorporating all four layers (tail dependence, continents, sectors, and global) consistently yielded the highest predictive accuracy for both the 2008 and 2020 financial crises.
  • PMFG Layer's Significance: The tail dependence (PMFG) layer, when combined with the global layer, provided the second-best predictive performance, underscoring its importance in capturing fundamental financial linkages.
  • Dynamic Prediction: The model's ability to predict future infections was evaluated using a sliding window approach, showing that optimal prediction windows (n days of past data) vary between crises but generally improve with more data, up to a certain point.
  • Geographic and Sectoral Insights: The model successfully predicted the geographical regions and economic sectors most likely to be affected by a crisis, with the Sørensen-Dice coefficient indicating good similarity between predicted and actual outcomes.

Limitations and Future Work

The authors acknowledge several limitations:

  • Reliance on Stock Price Data: The model's predictions are primarily based on stock price volatility, which offers an incomplete picture of a company's overall financial stability.
  • Data Availability and Quality: The accuracy of predictions is contingent on the availability and quality of historical data, which can vary significantly across companies, sectors, and regions.
  • Model Assumptions: The SIR model's division into susceptible, infected, and recovered states simplifies reality and may not fully capture the nuances of financial contagion.
  • Omitted Factors: The model does not explicitly account for other crucial factors like macroeconomic indicators, regulatory policies, investor sentiment, or systemic risk events.

Future work suggested includes:

  • Dynamic Network Updates: Incorporating dynamic updates to the multiplex connectivity structure and model parameters to better capture periods of rapid change.
  • Directed Graphs: Exploring the impact of using directed rather than undirected graphs to represent transmission pathways.
  • Additional Information: Integrating external information, such as government stimulus packages or lockdown periods, to improve parameter updates during significant change points.
  • Company-Specific Predictions: Further detailed analysis of the model's accuracy in predicting which specific companies are likely to be affected.

Summary updated

Abstract

This paper proposes a novel framework for modelling the spread of financial crises in complex networks, combining financial data, Extreme Value Theory and an epidemiological transmission model. We accommodate two key aspects of contagion modelling: fundamentals-based contagion, where the transmission is due to direct financial linkages, and pure contagion, where a crisis might trigger additional crises due to global effects. We use stock price, geographical location and economic sector data for a set of 398 companies to construct multiplex networks of four layers, on which a susceptible-infected-recovered transmission model is defined, in order to model the spread of financial shocks between companies by accounting for their interconnected nature. By utilizing stock price data for the 2008 and 2020 financial crises, we investigate and assess the effectiveness of our model in forecasting the propagation of financial shocks through the network, where a shock is detected by measuring stock price volatility. The results suggest that the proposed framework is effective in predicting the spread of financial crises. Our findings demonstrate the significance of each layer of the multiplex network structure, which differentiates between various transmission pathways, for predicting the number of affected companies, as well as for company-, sector- or location-specific predictions.
BibTeX
@inproceedings{bozhidarova2024describing,
  author = {Malvina Bozhidarova and Frank Ball and Yves van Gennip and Reuben D. O’Dea and Gilles Stupfler},
  title = {{Describing financial crisis propagation through epidemic modelling on multiplex networks}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2024},
  doi = {10.1098/rspa.2023.0787},
}
Figure