Agathe Sadeghi, Ph.D. Candidate in Financial Engineering

Bio

Agathe Sadeghi is a Ph.D. candidate in Financial Engineering supervised by Dr. Zachary Feinstein. Her research focuses on network theory and risk factors, with applications in trading and portfolio management. She also explores causal discovery and inference to identify potential drivers of different asset classes. She is an instructor at Stevens, where she teaches the financial database course to graduate students.

Skillset

In addition to her expertise in statistical analysis, time series modeling, and machine learning, Agathe has strong presentation skills and the ability to effectively communicate complex ideas to diverse audiences, as well as translate business requirements into technical solutions.

Dissertation Summary

Network-based contagion risk factor for Joint VaR and ES Portfolio Optimization

In the Thesis, we introduce a novel contagion centrality measure called the Leontief Centrality Measure (LCM) to assess the systemic risk and shock propagation in financial networks. Unlike traditional centrality metrics, LCM is built on an absolute scale, allowing for meaningful comparisons across different time periods. We emphasize the need for statistical validation of network-based risk measures and provides an innovative testing framework to assess the accuracy of contagion risks through deriving the distribution of the measure and hypothesis testing. In empirical case studies using financial data, such as Credit Default Swap (CDX) indices and equity tick data, we show that the LCM is highly responsive during periods of financial distress, capturing underlying market dynamics and offering clearer insights than existing centrality measures like Degree Centrality and DebtRank. This framework has direct implications for identifying systemic risks and guiding investment strategies during turbulent market conditions.

We extend the application of the LCM measure by incorporating it as a risk factor to predict Valueat Risk (VaR) and Expected Shortfall (ES). The study compares several models from the literature, including CAViaR, GAS and linear Asymmetric Slope models, all developed using a neural network architecture. These models use lagged returns and the LCM measure calculated from equity tickdata of 20 financial companies, spanning from 2009 to mid-2024, to estimate the VaR and ES for the XLF index. We find that incorporating LCM into these models significantly improves the accuracy of the risk measures’ predictions compared to models that do not include LCM. The performance is evaluated using various loss functions—quantile loss and BCGNS for separate VaR and ESestimations, and FZ loss for joint predictions, which is crucial since ES is not directly elicitable. The study highlights that jointly predicting VaR and ES using LCM leads to better trading positioning and portfolio risk management.

The next step is to increase the complexity of the underlying network, enabling a multilayer network analysis that can capture interactions across multiple markets simultaneously, offering deeper insights into their dynamics. Ultimately, this approach could be leveraged for cross-market investment strategies, providing a clearer understanding of how assets in different markets are interconnected. We also plan to incorporate fintech elements, such as crypto markets, to evaluate the model's effectiveness in decentralized finance (DeFi) environments and explore its potential for cross-market portfolio diversification.