JDAPP: Reinforcement learning with phaseshift jump diffusion and Fourier randomization for robust financial agent training

We propose a reinforcement learning framework that enhances trading robustness through advanced synthetic data generation. The method integrates APARCH-based volatility modeling, jump-diffusion dynamics, and a novel Fourier-Block Randomization technique, termed JDAPP, to create diverse, non-repetitive return paths. A Double Deep Q-Network (DDQN) agent trained solely on these scenarios is evaluated against real S&P 500 data. Even with a binary long-or-short regime, the agent significantly outperforms a random baseline. Beyond performance, JDAPP enables stress testing of strategies under extreme market conditions, supporting risk management, regulatory oversight, and the development of more robust AI-based trading systems. • Introduces JDAPP, a reproducible jump–diffusion and APARCH synthetic data framework. • Combines volatility modeling, Fourier block randomization, and reinforcement learning. • Captures heavy tails, volatility clustering, and jump dynamics observed in SPY data. • Trained agents achieve higher Sharpe ratios and lower drawdowns than random baselines. • Provides full code-aligned appendices ensuring transparency and replicability.