GARCH-UGH: A bias-reduced approach for dynamic extreme Value-at-Risk estimation in financial time series

Hibiki Kaibuchi, Yoshinori Kawasaki, Gilles Stupfler

Introduction

In the realm of financial risk management, Value-at-Risk (VaR) serves as a fundamental metric for quantifying potential losses within a specific timeframe and confidence level. However, estimating VaR at extreme levels (e.g., 99.9%) presents significant challenges. Financial time series are characterized by two primary hurdles: volatility clustering, where periods of high or low volatility tend to persist, and heavy tails, where extreme market crashes occur more frequently than predicted by a normal distribution.

Traditional approaches, such as the GARCH-EVT framework, attempt to address these issues by first filtering the data for volatility and then applying Extreme Value Theory (EVT) to the residuals. While theoretically sound, these methods often suffer from high bias in tail estimation or rely on computationally intensive Maximum Likelihood Estimation (MLE) that can be difficult to calibrate. This paper introduces GARCH-UGH (Unbiased Gomes-de Haan), a two-step framework designed to provide more accurate and stable risk forecasts by explicitly correcting for estimation bias.

The GARCH-UGH Framework

The GARCH-UGH methodology decomposes the estimation process into two distinct stages to handle both the dynamic nature of market volatility and the statistical nuances of extreme tails.

Step 1: AR-GARCH Filtering

The process begins with an AR(1)-GARCH(1,1) model used to "pre-whiten" the return data. This step captures the time-varying mean and conditional variance, effectively filtering out the day-to-day fluctuations in market turbulence. The resulting standardized residuals are ideally independent and identically distributed (i.i.d.), providing a cleaner dataset for tail analysis.

Step 2: Bias-Reduced Tail Estimation

The core innovation of this research lies in the application of the Unbiased Gomes-de Haan (UGH) estimator to the filtered residuals. Traditional semi-parametric tail estimators, such as the Hill estimator, are highly sensitive to the choice of the threshold parameter $k$ (the number of upper order statistics used). If $k$ is too small, the estimator suffers from high variance; if $k$ is too large, bias from the central part of the distribution distorts the results.

The UGH estimator addresses this by utilizing a "second-order condition" to explicitly estimate and subtract the inherent bias. This makes the final VaR estimate significantly more robust to the choice of $k$, allowing risk managers to obtain reliable forecasts without the need for precise tuning of the tail threshold.

Empirical Evaluation

The model was evaluated using approximately 16 years of daily observations (4,000 data points) across four major financial indices, covering the volatile period surrounding the 2008 financial crisis:

  • Dow Jones Industrial Average (DJ)
  • Nasdaq Stock Market Index (NASDAQ)
  • Nikkei 225 (NIKKEI)
  • JPY/GBP Exchange Rate

Key Results and Backtesting

The researchers employed rigorous backtesting—comparing predicted risk to actual historical outcomes—to validate the model's performance. The results demonstrated several key advantages:

  1. Superior Accuracy: In-sample backtesting showed that GARCH-UGH was the most successful model in hitting target VaR violation counts in 46 out of 60 tested cases. In contrast, conventional GARCH-EVT models frequently underperformed.
  2. Regulatory Compliance: GARCH-UGH consistently passed standard regulatory assessments, including the Kupiec (frequency) and Christoffersen (independence) tests, for the Dow Jones and Nikkei indices. Standard methods failed these tests multiple times during the same period.
  3. Stability Across Thresholds: One of the most striking findings was the stability of GARCH-UGH estimates across a wide range of sample fractions ($k$ ranging from 5% to 25%). While standard methods saw accuracy degrade as more data was included from the "bulk" of the distribution, the bias-reduction in GARCH-UGH maintained precision.
  4. Computational Efficiency: Unlike MLE-based approaches that may face convergence issues, GARCH-UGH utilizes closed-form-like estimators. This makes the framework computationally efficient and easier to implement in high-frequency production environments.

Conclusion and Future Directions

GARCH-UGH represents a significant advancement in dynamic extreme risk estimation. By combining traditional volatility filtering with a bias-reduced semi-parametric estimator, the framework offers a robust solution to the bias-variance tradeoff inherent in tail estimation.

While the current research utilizes a standard AR(1)-GARCH(1,1) filter, future iterations could explore more sophisticated filters, such as asymmetric GARCH models, to further improve the pre-whitening stage. Additionally, extending this bias-reduction technique to other risk measures like Expected Shortfall (which quantifies the average loss beyond the VaR threshold) and adapting the model for multi-day forecasting horizons remain promising areas for future study.

Summary updated

Abstract

The Value-at-Risk (VaR) is a widely used instrument in financial risk management. The question of estimating the VaR of loss return distributions at extreme levels is an important question in financial applications, both from operational and regulatory perspectives; in particular, the dynamic estimation of extreme VaR given the recent past has received substantial attention. We propose here a two-step bias-reduced estimation methodology called GARCH-UGH (Unbiased Gomes-de Haan), whereby financial returns are first filtered using an AR-GARCH model, and then a bias-reduced estimator of extreme quantiles is applied to the standardized residuals to estimate one-step ahead dynamic extreme VaR. Our results indicate that the GARCH-UGH estimates are more accurate than those obtained by combining conventional AR-GARCH filtering and extreme value estimates from the perspective of in-sample and out-of-sample backtestings of historical daily returns on several financial time series.
BibTeX
@inproceedings{kaibuchi2022garch,
  author = {Hibiki Kaibuchi and Yoshinori Kawasaki and Gilles Stupfler},
  title = {{GARCH-UGH: A bias-reduced approach for dynamic extreme Value-at-Risk estimation in financial time series}},
  booktitle = {Quantitative},
  year = {2022},
  doi = {10.1080/14697688.2022.2048061},
}
Figure