What it does

This is a tutorial-grade stock trading bot built on Deep Q-Learning — the same family of algorithms that learned to play Breakout. It observes an n-day window of normalized price differences, then decides to buy, sell, or hold one share at a time. The reward signal is simply the change in your portfolio value. Training and evaluation scripts are included, plus a Jupyter notebook for visualizing trades.

The interesting bit

The author explicitly keeps the code “as close as possible to the paper, for learning purposes.” That honesty is refreshing: this is pedagogy with a portfolio, not a production system. It also implements three DQN variants — vanilla, fixed target, and Double DQN — while leaving Prioritized Experience Replay and Dueling Networks as unchecked homework.

Key highlights

• Trained on GOOG 2010–2017, validated on 2018 ($863.41 profit), tested on 2019 ($1,141.45 profit)
• Three DQN strategies implemented: dqn, t-dqn, and double-dqn
• Feature engineering is deliberately simple: sigmoid-normalized differences in adjusted closing price
• CPU training is preferred due to sequential episode structure
• Includes sample data and a visualization notebook

Caveats

• Single-share trades only; position sizing is explicitly out of scope
• Results shown for one stock over a specific bull-market period; generalization is unclear
• The README contains typos (“abserves,” “prefferably”) that suggest limited maintenance

Verdict

Worth a weekend if you’re trying to bridge the gap between DQN papers and something tangible. Skip it if you need risk management, multi-asset portfolios, or any illusion that past profits predict future performance.