Using quantile time series and historical simulation to forecast financial risk multiple steps ahead
Richard Gerlach, Antonio Naimoli, Giuseppe Storti · 2025
Read Using quantile time series and historical simulation to forecast financial risk multiple steps ahead by Richard Gerlach, Antonio Naimoli, Giuseppe Storti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-20978.txt.
Introduction
Richard Gerlach, Antonio Naimoli, and Giuseppe Storti develop a method for forecasting financial tail risk several steps ahead. Their targets are Value-at-Risk and Expected Shortfall, the standard measures of extreme losses. The approach is quantile-based and semi-parametric: returns are scaled by an estimated quantile series, then resampled to build forecast distributions at multiple horizons.
The method works whenever the relation between the two risk measures is stable over time. It extends to models incorporating realized volatility measures, in the spirit of realized GARCH and CAViaR frameworks. Simulations check finite-sample behavior, and a forecasting study compares one-day and ten-day forecasts at the one and two-and-a-half percent levels against competitors.
Catalog record
| CREATOR | Richard Gerlach, Antonio Naimoli, Giuseppe Storti |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 77 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-20978
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Using quantile time series and historical simulation to forecast financial risk multiple steps ahead Richard Gerlach, Antonio Naimoli, Giuseppe Storti Abstract A method for quantile-based, semi-parametric historical simulation estimation of multiple step ahead Value-at-Risk (VaR) and Expected Shortfall (ES) models is developed. It uses the quantile loss function, analogous to how the quasi-likelihood is employed by standard historical simulation methods. The returns data are scaled by the estimated quantile series, then resampling is employed to estimate the forecast distribution one and multiple steps ahead, allowing tail risk forecasting. The proposed method is applicable to any data or model where the relationship between VaR and ES does not change over time and can be extended to allow a measurement equation incorporating realized measures, thus including Realized GARCH and Realized CAViaR type models. Its finite sample properties, and its comparison with existing historical simulation methods, are evaluated via a simulation study. A forecasting study assesses the relative accuracy of the 1% and 2.5% VaR and ES one-day-ahead and ten-day-ahead forecasting results for the proposed class of models compared to several competitors. Keywords : Quantile Regression, Historical simulation, Value-at-Risk, Expected Shortfall, Multiple step ahead forecasting. JEL codes : C22, C53, C58 1 INTRODUCTION Since its introduction by J.P. Morgan in 1993, Value-at-Risk (VaR) is widely employed by financial institutions for capital allocation and risk management purposes. Let \mathcal{I}_{t} be the information available at time t and F_{t}(r)=Pr(r_{t}\leq r|\mathcal{I}_{t-1}) be the Cumulative Distribution Function (CDF) of the return r_{t} conditional on \mathcal{I}_{t-1} . Assuming that F_{t}(.) is strictly increasing and continuous on the real line {\rm I\!R} , VaR is the negative of the one-step-ahead \alpha level quantile at time t , defined as: Q_{\alpha,t}=F^{-1}_{t}(\alpha),\qquad 0<\alpha<1. VaR cannot measure typical loss for violations and is not mathematically coherent, i.e., it can favor non-diversification. Expected Shortfall (ES), proposed by Artzner 1997 and Artzner et al. 1999 , gives the expected loss, conditional on returns exceeding a VaR threshold, and is a coherent risk measure; it is also widely employed for tail risk measurement and is now favored by the Basel Committee on Banking Supervision. Under the same assumptions, the one-step-ahead \alpha level ES can be shown (see Acerbi and Tasche 2002 , among others) to equal the negative of the tail conditional expectation of r_{t} , i.e.: \text{ES}_{\alpha,t}=-E(r_{t}|r_{t}\leq Q_{\alpha;t},\mathcal{I}_{t-1}). Historical simulation (HS) for tail risk forecasting is well known. However, standard HS methods suffer from theoretical shortcomings that limit their ability to accurately predict risk under realistic dependence regimes, such as volatility clustering. A popular extension of the HS method uses quasi-maximum likelihood (QML) estimation of GARCH-type volatility models to first filter the return data, while applying the HS method only in a second step to estimate the quantile of the QML standardized residuals, see for example Gao and Song 2008 . This is referred to as filtered HS (FHS). This paper extends the FHS idea by considering filtering via conditional quantile models, to be denoted as Quantile FHS (QFHS). Manganelli and Engle 2004 develop a method that we consider a pre-cursor to our proposed QFHS methods. They employed standard CAViaR type models ( Engle and Manganelli 2004 ) , estimated by minimising the quantile loss function, to estimate VaR; then regressed the returns exceeding the estimated VaR series on the VaR series itself, to estimate ES ( Manganelli and Engle 2004 , see Equations 9.23-9.25 in) . This method is equivalent to our QFHS approach for h=1 period forecasts, but cannot be employed for multi-step ahead forecasting. More recently, relying on the theoretical results on the elicitability of the pair (VaR, ES) derived by Fissler and Ziegel 2016 , Taylor 2019 , Patton et al. 2019 , Mitrodima and Oberoi 2023 and Chen et al. 2023 have proposed joint dynamic semi-parametric models for VaR and ES, often called ES-CAViaR models. The last two contributions also allow for the modelling of long range dependence in VaR and ES dynamics. However, none of these models, as with the original CAViaR model, can be used to generate multi-step-ahead risk forecasts. This paper develops and extends both the QML-based FHS method and the aforementioned method in Manganelli and Engle 2004 in three different directions. First, we illustrate that regression-based ES estimation is equivalent to a HS method, based on scaling returns by the estimated VaR series. Then, we extend the method so that VaR and ES forecasting can be done at any quantile level, independent and separate to the choice of risk level used for estimating the CAViaR model. Simulations are conducted and presented here, to investigate whether there are more optimal quantile levels for estimation, and how this quantile loss based estimation compares to the standard quasi-likelihood estimation in existing FHS methods, e.g. see Gao and Song 2008 and Francq and Zakoïan 2015 . Third, we further extend our proposed QFHS method, employing resampling to allow multi-step ahead forecasting for VaR and ES, via sampling from the full, implied multi-step ahead return distribution. This is an important feature, especially considering the focus of the Basel Accord on 10 day ahead risk forecasting. We illustrate how realized measures can be incorporated into the proposed method, by employing the pseudo-likelihood developed by Gerlach and Wang 2020 and Peiris et al. 2024 in their (log) Realized CAViaR model(s). Finally, we show that the proposed method implicitly assumes, and is applicable to, any data or model where the relationship between VaR and ES in the conditional return distribution does not change over time; which includes the case where the conditional return distribution is constant. Thus, the proposed method can be applied to most modern financial time series models and return datasets. In summary, the main contribution of this paper is the proposal of a new class of quantile-based filtered HS methods for estimating the one-step-ahead and multi-step-ahead conditional return distribution, for application to VaR and ES forecasting. The finite sample properties of this method are assessed via a simulation study and compared to existing FHS approaches. In the empirical study, the proposed methods are assessed via their VaR and ES forecasting performance and compared to a range of existing methods, illustrating generally favourable results. The paper is organized as follows. Section 2 briefly provides some background, reviewing existing methods related to our proposed method. Section 3 proposes and describes the new methods, including the associated quantile loss and quasi-likelihood employed for parameter estimation. Section 4 summarizes the simulations conducted. Section 5 presents the forecasting study. Section 6 concludes the paper and discusses future work. 2 BACKGROUND HS for tail risk forecasting is well known, though standard methods tend to not accurately predict risk under volatility clustering. The most successful extension of the HS method uses QML estimation of GARCH-type volatility models to first filter the return data, while applying the HS method only in a second step to estimate the quantile of the QML standardized residuals, see for example Gao and Song 2008 . This is known as filtered HS (FHS). The main steps of the FHS estimation procedure are briefly recalled below. Consider a 0-mean GARCH(1,1) process: \displaystyle r_{t} \displaystyle= \displaystyle\sigma_{t}\epsilon_{t}\,\,;\,\,\epsilon_{t}\sim\text{i.i.d.}\,D(0,1) (1) \displaystyle\sigma_{t}^{2} \displaystyle= \displaystyle\omega+\gamma r_{t-1}^{2}+\beta\sigma^{2}_{t-1} (2) where the CDF of \epsilon_{t} (i.e. of D(\mu,\sigma) ), is assumed to be