Deep Reinforcement Learning Versus MPT

Does machine learning reliably offer better risk-adjusted portfolio performance than traditional modern portfolio theory (MPT)? In their August 2023 paper entitled “Comparing Deep RL and Traditional Financial Portfolio Methods”, Eric Benhamou, Jean-Jacques Ohana, Beatrice Guez, David Saltiel, Rida Laraki and Jamal Atif compare principles, methodologies and risk-adjusted performances of dynamic deep reinforcement learning (DRL) and MPT. The DRL approach seeks long-only allocations that maximize Sharpe ratio (calculated assuming a zero risk-free rate). DRL training data includes individual asset returns, portfolio drawdown and contextual variables including U.S. and European interest rates, the CBOE volatility index (VIX), credit default swap prices, currency rates (U.S. dollar index), GDP and CPI forecasts, crude oil/gold/copper inventories and global, U.S., European, Japanese and emerging markets economic surprise indexes. DRL training employs an expanding window, each year training on available historical data and testing on the next year. They consider three MPT portfolios also using expanding window of historical data to estimate inputs: (1) full MPT (Markowitz); (2) minimum variance; and, (3) risk parity. Their global test data consists of daily returns of 11 futures contract series for four major equity indexes, four major bond indexes and three major commodity indexes. They assume trading frictions of 0.02% of value traded. Using the specified (groomed) data during 2000 through mid-2023, they find that: (more…)