I am currently trying to do a multivariate forecast using the DCCroll function for a portfolio of 15 stocks (Forecast length should be 400). I tried it two times (reduced the forecast length in the second run): After around 200-260 times the function broke
down due to the fact that the "system is computationally singular". I understand that this results from linearly dependent columns , i.e. strongly correlated variables or more variables than observations (I have 1216 observations per stock and 15 stocks. So
this should not be the problem, I assume.)
When I checked my return correlationmatrix and my realized volatilities correlation matrix, there were no indications that I face issues such as strong correlation.
Stock Return Correlation Matrix:
Stock RV Correlation Matrix:
Also what seems odd to me is that the forecast can be done for the first 200-260 days and then it breaks down, as this suggests that the majority of the forecasting process is working.
Further I already forecast the univariate realGARCH time series with the same stocks before. So I know that the first stage of the DCC estimation should work.
What are possible solution strategies that I can follow?