Abstract:
This paper explores the possible advantages of introducing observable state variables into risk
management models as a strategy for modeling the evolution of second moments. A simulation
exercise demonstrates that if asset returns depend upon a set of underlying state variables that are
autoregressively conditionally heteroskedastic (ARCH), then a risk management model that fails to
take account of this dependence can badly mismeasure a portfolio's "Value-at-Risk" (VaR), even if the
model allows for conditional heteroskedasticity in asset returns. Variables measuring macroeconomic
news are constructed as the orthogonalized residuals from a vector autoregression (VAR). These news
variables are found to have some explanatory power for asset returns. We also estimate a model of
asset returns in which time variation in variances and covariances derives only from conditional
heteroskedasticity in the underlying macroeconomic shocks. Although the data give some support for
several of the specifications that we tried, neither these models nor GARCH models that used only
asset returns appear to have much ability to forecast the second moments of returns. Finally, we allow
asset return variances and covariances to depend directly on unemployment rates -- proxying for the
general state of the economy -- and find fairly strong evidence for this sort of specification relative to
a null hypothesis of homoskedasticity.
|
