Dynamic Rolling Out-of-Sample Portfolio Optimization, Covariance Shrinkage, and L2 Regularization in Indian Equity Markets

This paper evaluates modern portfolio optimization across an expanded universe of 20 blue-chip constituents from the NIFTY 50 index across the 2019–2026 period. To resolve classical Markowitz failure modes—specifically input parameter noise, the "Error Maximizer" paradox, and corner-solution over-concentration—we introduce a robust quantitative framework combining analytical Ledoit-Wolf linear covariance shrinkage, L2 Ridge regularization, hard weight bounds (0% ≤ w_i ≤ 15%), turnover transaction friction penalties (15 bps), and dynamic monthly rolling-window rebalancing (36-month lookback). Evaluated on unseen out-of-sample data (2022–2026), the regularized shrunk optimal portfolio achieves an annualized net return of 8.95% with an annualized volatility of 14.11% (Sharpe Ratio = 0.1598 at R_f = 6.7%), compared to 9.14% annualized net return and 12.81% volatility (Sharpe Ratio = 0.1902) for the naive 1/N baseline. Historical 95% Value-at-Risk (VaR) analysis demonstrates controlled daily tail loss (1.36% vs 1.26%). We analyze the empirical limits of mean-variance optimization against naive diversification (1/N) and document how covariance shrinkage suppresses parameter noise during dynamic rebalancing regimes.