MCDM-Guided Deep Reinforcement Learning for Dynamic Portfolio Optimization

Abstract

When there is uncertainty, portfolio optimisation is the process of figuring out how to invest money in a group of assets. This is an important part of managing assets. Intellectual portfolio optimisation is now necessary in today's financial markets because the market is becoming more complicated and unstable. Current portfolio optimisation methods are either static or based solely on data-driven learning, which cannot account for the generalised multiple decision metrics, joint investor state, and time correlation. This model learns from a professionally balanced dataset that includes historical stock price data from Yahoo Finance, macroeconomic indices from the Federal Reserve Economic Data database (FRED), and systematic risk factors from the Fama–French data library.The framework uses both MCSMs and deep reinforcement learning (DRL) to manage portfolios that change over time. MCDM ranks the assets using a number of financial and economic measures. A deep reinforcement learning agent will also keep track of the portfolio weights over time, but only in a small action space. In this study, we innovatively employ reinforcement learning alongside MCDM-based asset prioritisations to improve stability, interpretability, and risk evaluation. We use standard risk-adjusted metrics like cumulative return, Sharpe and Sortino ratios, and maximum drawdown to judge how well a portfolio is doing. We conduct numerical experiments to verify that MCDM-guided RL significantly outperforms portfolio benchmarks in terms of risk-adjusted return and maximum drawdown. This shows that this method works for making portfolios smarter.