FHS with DCC

Interactive exploration of Filtered Historical Simulation combined with Dynamic Conditional Correlations for portfolio risk with time-varying dependence

When correlations are dynamic, the constant-correlation multivariate FHS must be extended. Historical shocks were correlated according to their historical (time-varying) correlations, so to use them in forward simulation with different current correlations, we must first uncorrelate them and then re-correlate with the simulated dynamic correlations (Christoffersen 2012, chap. 8; Engle 2002).

Recall that the covariance matrix decomposes as \(\Sigma_{t+1} = D_{t+1}\,\Upsilon_{t+1}\,D_{t+1}\), where \(D_{t+1}\) is the diagonal matrix of GARCH standard deviations and \(\Upsilon_{t+1}\) is the DCC correlation matrix. \(\Upsilon^{1/2}\) denotes its Cholesky decomposition (the matrix square root used to correlate shocks).

The DCC-FHS procedure:

  1. Uncorrelate historical shocks: \(\hat{z}^u_{t+1-\tau} = \Upsilon^{-1/2}_{t+1-\tau}\hat{z}_{t+1-\tau}\)
  2. Draw uncorrelated shock vectors (entire vector from the same day)
  3. Re-correlate with current correlations: \(\hat{z}_{i,t+1} = \Upsilon^{1/2}_{t+1}\hat{z}^u_{i,1}\)
  4. Compute returns: \(\hat{r}_{i,t+1} = D_{t+1}\hat{z}_{i,t+1}\)
  5. Update both GARCH variances and DCC correlations using simulated shocks
  6. Repeat for subsequent days \(k = 2, \ldots, K\)
Note

Why uncorrelate? If we drew correlated historical shocks directly and applied them with the current correlation matrix, the shocks would be “double-correlated”: once from their historical correlation structure and once from the re-correlation step. Uncorrelating first removes the historical dependence, creating a clean database from which to re-impose the current (and dynamically evolving) correlation.

The uncorrelate/re-correlate pipeline

This illustration shows the three stages of the DCC-FHS shock transformation and compares portfolio risk under constant vs. dynamic correlations.

Tip

How to experiment

Set the current conditional correlation \(\rho_{12,t+1}\) to a high “crisis” value (e.g., 0.80) and compare with a “normal” value (0.30) in the stress scenario tab. This is the one-day-ahead correlation forecast from the DCC model at the end of day \(t\), analogous to the GARCH volatility forecast \(\sigma_{i,t+1}\). Increase the simulation horizon to see how the DCC correlation mean-reverts toward \(\bar{\rho}\), reducing the gap between the constant and dynamic approaches. Try different DCC parameters: higher \(\alpha\) makes the correlation more reactive to shocks; higher \(\beta\) makes it more persistent.

References

Christoffersen, Peter F. 2012. Elements of Financial Risk Management. 2nd ed. Academic Press.
Engle, Robert. 2002. “Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive Conditional Heteroskedasticity Models.” Journal of Business & Economic Statistics 20 (3): 339–50.