Stationary Time Series: Practical Implications for Risk Modeling
A risk analyst at a global investment bank is assessing the daily changes in a portfolio's 5-year corporate bond credit spread. Over the past year, the unconditional mean of these daily changes has been statistically indistinguishable from zero. However, Ljung-Box tests on the squared daily changes indicate significant autocorrelation at multiple lags, and a review of the 60-day rolling standard deviation reveals distinct periods of both elevated and subdued volatility. The current 1-day 99% VaR for this portfolio component is calculated using a 250-day equally-weighted historical simulation, which implicitly assumes independently and identically distributed (i.i.d.) returns. What is the most significant practical implication for the existing VaR model, and what immediate analytical adjustment is most appropriate?
A.The series exhibits non-stationary variance due to volatility clustering, violating the i.i.d. assumption for the historical simulation, necessitating a time-varying volatility model like GARCH to accurately capture risk.
B.The daily changes are indicative of a unit root process, requiring further differencing to achieve mean stationarity before any valid statistical inference can be made regarding the VaR.
C.The non-stationarity observed in the changes implies the original credit spread levels are not mean-reverting, making any VaR calculation unreliable unless a cointegration analysis is first performed.
D.The historical simulation window should be substantially extended (e.g., to 750 days) to encompass a wider range of volatility regimes and improve the unconditional variance estimate for the VaR.
Rationale:
The description clearly indicates that while the mean of the daily changes is stable (statistically indistinguishable from zero), the variance is not. Significant autocorrelation in squared daily changes and varying rolling standard deviation are classic signs of heteroskedasticity and volatility clustering. This means the series is not covariance stationary, violating the independently and identically distributed (i.i.d.) assumption underlying standard historical simulation VaR. This violation makes the VaR estimate unreliable, as it does not adequately capture the dynamic nature of volatility. The most appropriate immediate adjustment is to consider models specifically designed for time-varying volatility, such as Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models, which can forecast conditional variances more accurately.
The idea of a unit root process and further differencing is incorrect because the problem states the mean of the daily changes is statistically indistinguishable from zero, implying mean stationarity has likely been achieved for the first-differenced series. Applying further differencing would likely lead to overdifferencing and introduce spurious moving average components.
The claim about original credit spread levels not being mean-reverting and requiring cointegration is misdirected. The core issue presented is with the *variance* characteristics of the *daily changes*, not the long-term mean-reversion properties of the *levels* or relationships between multiple series, which is what cointegration addresses. While levels might be non-stationary, the problem's focus is on the suitability of the VaR model for the *changes* given their observed properties.
The suggestion to extend the historical simulation window is an inadequate solution. While a longer window might encompass more volatility regimes in aggregate, it still treats all observations equally within the window, failing to capture the *dynamics* of volatility clustering and how recent volatility might be more indicative of current risk. It does not address the fundamental issue of time-varying volatility and can smooth out critical, recent risk concentrations.
100% Free
Free FRM Questions + Study Tips
Join FRM candidates receiving practice questions and expert exam-day strategies directly to their inbox.
