The black - litterman model estimates the expected return using the bavesian shrinkage model (historical samples covariance matrix plus your prior view). Therefore, if there were no views regard one of the considered assets, it’s only possible to use its historical data base, which gives unstable results.I was wondering if it would be consistent to use the Bengstsson and Holst (2002) model (or any other “shrinkage” model) in order to obtain a more robust matrix and use it to feed the back-litterman model? Because I suspect that using any matrix creators models it can brings some inconsistency, due to the data base overlap when using the views over the assets (shrinkage of the shrinkage?)Does anybody already test it or uses those models? If yes, How do you work with the covariance matrix?