Dynamic Modeling and Risk Constraint Analysis for Multi-Asset Portfolio Optimization

Abstract

With increasing volatility in global financial markets, traditional static portfolio optimization models struggle to adapt to the time-varying characteristics of asset returns and risks. To address the dynamic optimization needs of multi-asset portfolios, this paper proposes a dynamic robust multi-asset portfolio optimization algorithm (DRMPOA). This algorithm first integrates macroeconomic indicators through an improved attention-based LSTM model to predict asset returns and volatility. A robust optimization model with CVaR constraints is then constructed, combining a particle swarm optimization algorithm with dynamic inertia weights to solve for optimal weights. These weights are then adjusted in real time using a 60-trading-day sliding window. Experiments are conducted using data from multiple asset classes, including the CSI 300 Index, the CSI All Bond Index, the Nanhua Commodity Index, and Bitcoin, from 2018 to 2023. The algorithm is compared with MV, Static-CVaR, and Dynamic-MCVaR algorithms. Results show that DRMPOA achieved a cumulative return of 48.2% (8.1% annualized), a 9.5 percentage point increase over Dynamic-MCVaR. Its 95th-percentile daily CVaR was 1.2%, and its annualized volatility was 12.3%, 0.1 and 1.5 percentage points lower than Dynamic-MCVaR, respectively. Its annualized Sharpe ratio was 0.58 (a 28.9% increase), and its single solution took 12.5 seconds (3.3 seconds faster than Dynamic-MCVaR). This algorithm balances profitability, risk tolerance, and efficiency, providing an effective solution for multi-asset allocation.