Applying Reinforcement Learning to Portfolio Optimization

PtrPomorski · x · 2026-07-19

This paper introduces SciPhyRL (Scientific Physics-Informed Reinforcement Learning), a dynamic framework for large-scale institutional portfolio optimization.

The core approach frames the continuous-time optimization problem within an expanded state space, explicitly accounting for cumulative costs, and then uses offline historical data to learn distribution-aware policies. The authors project the intractable HJB equations onto observed trajectories, converting them into path-based Hamilton-Jacobi equations, which are fitted directly in a single offline solve using PINNs, avoiding traditional value/policy iteration. To accommodate short holding periods, the control variable shifts from continuous trading rates to discrete target holdings, while execution costs are assessed using a microstructure-inspired quadratic impact model.

Experiments were conducted on a 14-asset ETF universe using artificially constructed oracle signals. Results show that the learned Gibbs policy outperforms equal-weight and behavioral baselines both in-sample and out-of-sample, achieving a significantly higher Sharpe ratio while better controlling volatility and turnover.

Original post →

More from Research

Research channel →