Dynamic Asset Allocation with Direct Policy Optimization: Evidence from a US Public-Market Multi-Asset Universe

We consider an asset allocation problem where funds must be allocated to six broad classes of assets—equity, corporate bonds, government bonds, listed real estate, gold, and commodities—each represented by a liquid, predominantly US public-market index that can be tracked with ETFs. We use a parsimonious reinforcement learning model to solve it, capitalizing on the time separability of the problem and the fact that for a price-taking investor, the state of the world is independent of her prior actions; these two features collapse the multi-period problem into a sequence of one-period direct policy optimizations, so that the policy network can be trained like a supervised model, without value functions or policy-gradient rollouts. In a walk-forward backtest from 20 October 2011 through 5 June 2026, the four-asset strategy earns a gross total return of 374% and a Sharpe ratio of 0.77, against 235% and 0.62 for Markowitz tangency portfolio and 222% and 0.68 for an equally weighted portfolio. Net of trading costs of 10 basis points, these are 342%, 228%, and 220%, with Sharpe ratios of 0.73, 0.61, and 0.68. In the six-asset case, which adds government and corporate bonds, our total return is 367% gross and 336% net, against 203% and 139% gross (199% and 138% net) for the two benchmarks. The Sharpe ratio, 0.79 gross and 0.75 net, exceeds equal weighting but not the tangency portfolio (0.91 and 0.89), which concentrates in low-volatility bonds. The sign of the return advantage is the same in four of five pre-defined monetary-policy regimes, although the individual regime-level differences are not statistically significant; the advantage survives a broad set of further benchmarks (including mean-variance rules with the same risk aversion, mean–expected-shortfall and risk-parity portfolios), and disappears only at trading costs above 60 basis points per unit traded (72 against the tangency portfolio).