DQN Stock Timing with Windowed Price States and All-In Trades

Summary

This study describes a revised deep Q-network approach for timing individual Chinese stocks. The earlier version used daily OHLCV data and several common factors as its state. The revision instead takes a configurable window of recent closing prices and transforms consecutive price differences into a state representation. Its backtest also changes position handling: a buy signal commits all available capital, while a sell signal exits the entire holding.

The authors compare reported DQN returns with buy-and-hold returns for a small set of stocks spanning rising, falling, and volatile price paths. Results vary: the DQN performs better on most listed examples, but trails the benchmark on at least one. This limited sample does not establish general effectiveness. The article notes that the strategy remains open to improvement and exposes several training parameters, while leaving network architecture settings less configurable.

Key ideas

• The revised DQN state uses a configurable window of closing prices transformed through consecutive price changes.
• The revised backtest enters with all available capital and exits the full position on a sell signal.
• The reported comparison uses buy-and-hold returns as the benchmark across several individual stocks.
• Results are mixed, and the small tested sample does not establish broad performance.
• Several training parameters are adjustable, while network architecture options remain limited.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.