Hands-On Machine Learning for Algorithmic Trading

A basic trading game

To train the agent, we need to set up a simple game with a limited set of options, a relatively low-dimensional state, and other parameters that can be easily modified and extended.

More specifically, the environment samples a stock price time series for a single ticker using a random start date to simulate a trading period that, by default, contains 252 days, or 1 year. The state contains the (scaled) price and volume, as well as some technical indicators like the percentile ranks of price and volume, a relative strength index (RSI), as well as 5- and 21-day returns. The agent can choose from three actions:

• Buy: Invest capital for a long position in the stock
• Flat: Hold cash only
• Sell short: Take a short position equal ...