This paper integrates recursive utility, a risk-sensitive objective from asset-pricing theory, into actor-critic reinforcement learning for portfolio allocation. The recursive utility framework separates risk aversion from intertemporal substitution, which is not captured by standard discounted sum of rewards. The authors approximate the certainty equivalent of future value using K-sample Monte Carlo and derive a Bellman residual-based advantage estimate (AAE) that generalizes to multi-step returns with state-dependent weights. The method is applicable only to critic-based algorithms like PPO and A2C, as it requires a learned value function.

Experiments on 10 chronological train/test splits of South Korean ETF data show that the recursive-utility agent improves upon the naive (discounted) baseline in Sharpe ratio (2.07 vs 1.22), max drawdown (10.38% vs 12.26%), and cumulative return (8.23% vs -6.47%). It also outperforms a Markowitz-based reward objective. However, the equal-weight (1/N) benchmark still achieves higher Sharpe ratios (2.3-2.7), indicating that the RL agents have room for improvement. The results are consistent with a risk-sensitive objective that penalizes downside dispersion, but high variance across splits and sensitivity to hyperparameters suggest caution in interpretation. The paper provides detailed derivations, experiment setup, and ablation studies in the appendices.