TL;DR
This paper introduces "expectHill" estimators, a novel approach to estimating extreme financial risks (like Expected Shortfall) for heavy-tailed distributions. By combining traditional order statistics with a more robust expectile-based method, these estimators offer improved accuracy and stability, particularly during financial crises.
What Problem Does This Paper Address?
Financial risk management heavily relies on quantifying potential losses. Traditional methods often use Value at Risk (VaR) and Expected Shortfall (ES), which are based on quantiles. However, quantiles have limitations: they only consider the frequency of extreme events, not their magnitude, and VaR, in particular, doesn't always satisfy desirable "coherence" properties for risk measures. Expectiles, a more recent alternative, address some of these shortcomings by considering both frequency and magnitude. The challenge this paper tackles is how to reliably estimate these expectile-based risk measures, especially in the extreme tails of heavy-tailed financial data, where observations are scarce and estimation is difficult.
Methods
The core innovation is the "expectHill" estimator for the tail index (a key parameter in heavy-tailed distributions). This estimator is a weighted combination of two components:
- Pure expectile-based estimators: These leverage the properties of expectiles to estimate the tail index.
- Hill estimator: A well-established estimator based on order statistics.
The paper then uses these expectHill estimators as a foundation to build estimators for:
- Tail expectiles: Estimating extreme values of expectiles.
- Extreme Expected Shortfall (XES): Estimating the expected loss given that the loss exceeds a very high threshold.
The authors develop the asymptotic theory for these estimators, providing mathematical guarantees for their behavior in large samples. They also propose a two-step procedure for selecting the optimal weights for their estimators, aiming to minimize asymptotic variance.
Datasets
The paper uses a real-world dataset of daily loss returns for the SPDR S&P 500 ETF (SPY) from June 15, 2004, to June 13, 2014. This dataset comprises 2,497 trading days. For robustness, they also analyze weekly loss returns for the same period, as well as specific sub-periods (Before Crisis, During Crisis, After Crisis) to examine the impact of financial events.
Key Insights and Ideas
- Expectiles offer a more robust alternative to quantiles: They capture both the frequency and magnitude of extreme events, leading to more stable risk measures.
- Combining expectiles and order statistics is powerful: The "expectHill" estimator leverages the strengths of both approaches for tail index estimation, leading to smoother and more stable estimates, especially during volatile periods.
- Financial crises highlight the importance of robust methods: The study shows that the difference between traditional (Hill) and expectHill estimators becomes more pronounced during crises, suggesting the latter's superiority in such conditions.
- A two-step weight selection strategy works well: The proposed method for choosing optimal weights for the expectHill estimators leads to results comparable to using theoretically optimal weights, offering a practical approach.
Key Results
- The expectHill estimators provide more stable and less volatile estimates of the tail index compared to the standard Hill estimator, particularly during the "During Crisis" period.
- The proposed estimators for extreme Expected Shortfall (XES) and quantile-based Expected Shortfall (QES) demonstrate good finite-sample performance in simulations across different heavy-tailed distributions (Student, Fréchet, Pareto).
- Analysis of the SPDR S&P 500 ETF data reveals that Expected Shortfall levels were significantly higher during the 2008-2009 crisis period, underscoring the need for robust risk estimation methods.
- The expectHill-based estimators for ES were found to be more efficient in the case of non-negative Fréchet and Pareto distributions, while other variants performed better for Student distributions.
Limitations and Future Work
- Bias: While the expectHill estimators are designed for low variance, their bias might be high, especially for certain parameter choices. Minimizing the Asymptotic Mean-Squared Error (AMSE) to address this is suggested as future work, though it requires estimating difficult second-order extreme value parameters.
- Serial Dependence: The current theory assumes independent and identically distributed data. The authors note that serial dependence in financial returns could affect asymptotic confidence intervals. Extending the theory to handle time series data (e.g., using β-mixing conditions) is a significant avenue for future research.
- Parameter Selection: While a practical two-step procedure for selecting weights is proposed, further investigation into optimal parameter selection for second-order extreme value parameters is needed.
Summary updated