Functional estimation of extreme conditional expectiles

Stéphane Girard, Gilles Stupfler, Antoine Usseglio‐CarleveEconometrics 2021 · DOI

TL;DR

This paper introduces a novel statistical framework for estimating extreme conditional expectiles in the presence of functional data. This is crucial for accurately assessing extreme risks in fields like finance and insurance when the risk depends on complex, high-dimensional predictor variables (like time series curves). The authors develop and analyze new estimators that can handle these challenging functional predictors.

What Problem Does This Paper Address?

In risk management, understanding extreme events is paramount. While quantiles (like Value-at-Risk) are common, they have limitations, particularly in not being "coherent risk measures" and not fully capturing the magnitude of extreme losses. Expectiles, which are based on tail expectations, offer a more robust alternative and are coherent risk measures. However, estimating these expectiles becomes significantly more complex when the outcome variable (e.g., financial loss) depends on a predictor that is not a simple number but a function (e.g., a time series of past market data). This paper tackles the problem of estimating extreme conditional expectiles – those at very high levels – when the predictor is a function, a scenario that has been largely unexplored.

Methods

The core of the paper lies in adapting existing techniques for estimating extreme conditional quantiles to the functional setting for expectiles. The authors leverage the interpretation of expectiles as quantiles of a modified distribution. Key methodological contributions include:

  • Functional Kernel Estimation: They employ a functional kernel estimator to approximate the conditional survival function of the outcome variable given the functional predictor. This is a standard technique in functional data analysis but applied here to a specific problem.
  • Expectile as a Quantile Transformation: A crucial insight is rewriting the conditional expectile as a quantile of a different, derived distribution. This allows them to use established quantile estimation machinery.
  • Weissman-Type Extrapolation: To estimate extreme expectiles (beyond the range of the observed data), they adapt the Weissman extrapolation technique. This involves estimating the conditional tail index (which describes how heavy the tail of the distribution is) and using it to extrapolate from intermediate expectile estimates to extreme ones.
  • Multiple Tail Index Estimators: The paper proposes and analyzes several estimators for the functional tail index, acknowledging its critical role in the extrapolation process.
  • Asymptotic Theory: A significant portion of the paper is dedicated to establishing the theoretical properties (asymptotic normality) of their proposed estimators under various assumptions, which is essential for understanding their reliability.

Datasets

The paper uses simulated data generated from three different stochastic processes for the functional covariate X:

  • (X1): A process based on a latent uniform random variable and a cosine function.
  • (X2): A process defined by a sum of scaled cosine functions with uniform random coefficients.
  • (X3): A standard Brownian bridge.

For real-world validation, they use a dataset of Bitcoin cryptocurrency hourly log-returns from January 1, 2012, to January 8, 2018, from the Bitstamp exchange. This dataset is structured to create pairs of (past hourly log-returns curve, next day's maximum hourly log-return).

Key Insights and Ideas

  • Expectiles as a Superior Risk Measure: The paper reinforces the idea that expectiles are a more suitable tool than quantiles for risk management due to their coherence properties and their ability to capture the magnitude of extreme events.
  • Functional Predictors Complicate Extreme Risk Estimation: The presence of functional predictors dramatically increases the complexity of estimating extreme risks, requiring specialized techniques beyond standard regression.
  • Tail Index is Key for Extrapolation: The accuracy of estimating extreme expectiles hinges critically on the accurate estimation of the conditional tail index, which quantifies the "heaviness" of the distribution's tail.
  • Bridging Theory and Practice: The paper successfully bridges theoretical statistical advancements with practical applications in finance and risk management, demonstrating the utility of their methods on real data.

Key Results

The paper presents several theoretical results regarding the asymptotic properties of their estimators. While specific numerical performance metrics are detailed in the simulation and real-data examples, the theoretical findings include:

  • Asymptotic Normality: The proposed estimators for intermediate and extreme conditional expectiles are shown to be asymptotically Gaussian under certain conditions, meaning their behavior can be well-approximated by a normal distribution for large sample sizes.
  • Performance Comparison: In simulations and real-data analysis (Figures 1, 3, 4, 6, 7, 8), the authors compare different estimators for the tail index and expectiles. They observe that their proposed estimators provide reasonable approximations, with some estimators showing advantages in specific scenarios (e.g., lower variance for certain tail index values or better performance with bias-reduced versions).
  • Real-World Applicability: The analysis on Bitcoin log-returns suggests that the proposed methods can capture the relationship between past volatility (represented by the functional predictor) and future extreme price movements, indicating their practical relevance.

Limitations and Future Work

  • Assumptions: The theoretical results rely on several assumptions, including the existence of finite conditional moments and specific tail behavior (regular variation). The practical implications of violating these assumptions are not fully explored.
  • Bandwidth and Parameter Selection: While cross-validation is used for bandwidth selection, choosing optimal parameters in functional extreme value analysis can still be challenging, especially in high-dimensional functional spaces.
  • Computational Complexity: Functional data analysis, particularly with kernel methods, can be computationally intensive. The paper does not extensively discuss the computational efficiency of their estimators.
  • Future Directions: The authors suggest exploring alternative semi-metrics for functional data, potentially leading to more efficient estimators. They also hint at using combinations of estimators (e.g., Pickands-type estimators) for tail index estimation, which could be a fruitful area for future research. The paper also mentions the potential for using functional Principal Component Analysis (FPCA) to simplify the functional predictor space, which could improve computational performance.

Summary updated

BibTeX
@inproceedings{girard2021functional,
  author = {Stéphane Girard and Gilles Stupfler and Antoine Usseglio‐Carleve},
  title = {{Functional estimation of extreme conditional expectiles}},
  booktitle = {Econometrics},
  year = {2021},
  doi = {10.1016/j.ecosta.2021.05.006},
}
Figure