Construction of Optimization Models and Empirical Testing for Multi-Asset Allocation in Quantitative Portfolio Management

Traditional multi-asset portfolios generally assume that historical correlations and diversification relationships are stable enough to be relied upon in optimization. That assumption fails during market stress, precisely when diversification matters most: correlations rise, historical hedges stop hedging, and static optimizers misallocate risk. This study develops and empirically evaluates a portfolio optimization framework that explicitly models changing cross-asset dependencies while accounting for the estimation uncertainty, liquidity heterogeneity, and trading costs that accompany implementation. Three practical problems motivate the design—correlations are estimated imperfectly, market regimes change, and trading is costly—and each is met by a corresponding safeguard: distributionally robust covariance estimation, regime-aware rebalancing triggers, and turnover regularization. The framework is tested on a balanced panel of 127 assets spanning equities, fixed income, commodities, and alternative risk premia over 365 weekly observations (2017–2023), with rolling out-of-sample evaluation across low-volatility, stress-transition, and high-inflation regimes, and is benchmarked against Global Minimum Variance (GMV) and equal-weight (EW) portfolios. Relative to those benchmarks, the framework reduces maximum drawdown depth during the 2020 Q1 and 2022 H2 stress episodes (12.3% versus 24.7% and 28.9%), lowers the tail risk ratio (1.82 versus 2.94 and 3.17), preserves a higher Effective Number of Bets (8.4 versus 4.1 and 5.9), and does so with lower turnover. Taken together, the results suggest that explicitly modeling time-varying cross-asset dependencies and implementation constraints, rather than optimizing against a static covariance matrix, can materially improve portfolio resilience; the comparison is limited to these two benchmarks and to a single historical sample, and is interpreted accordingly.