Can an agent learn to trade from experience alone? This project explores exactly that – a deep reinforcement-learning agent that learns a trading policy directly from historical price data, with no hand-coded rules.
It implements a Deep Q-Network that observes a normalized window of recent prices and chooses to sit, buy, or sell to maximize cumulative return. Training uses experience replay, an epsilon-greedy strategy that decays exploration from 1.0 to 0.01, and discounted rewards, delivered in both PyTorch and Keras/TensorFlow so the same value-based method can be benchmarked across frameworks.
It is a rigorous study in applying value-based RL to noisy financial time series, with saved models and reproducible training throughout – a research build, not financial advice.