Hedging American Put Options with Deep Reinforcement Learning: Training Procedures | HackerNoon

To train a Deep Reinforcement Learning (DRL) agent to effectively hedge American put options, robust asset price data generation and proper reward assignment based on option pricing are crucial.

Each DRL agent has the same architecture and hyperparameters, but the training procedure varies with two experiment rounds: one using a Geometric Brownian Motion (GBM) model and another utilizing a calibrated stochastic volatility approach.

For the stochastic volatility experiments, arbitrary model coefficients were first applied, and important parameters like the drift term and correlation were carefully calibrated to align with the actual market dynamics.

The approach ensures that DRL agents encounter scenarios below the expected exercise boundary, enhancing their learning and adaptability by exposing them to a range of market conditions.

Read at Hackernoon