RNN Models for High-Frequency Stock Price Direction

Summary

This article describes using recurrent neural networks to predict the direction of the next price change from high-frequency order-book data. The study uses data from Nasdaq stocks and keeps observations where prices change, reducing the input sequence. It applies truncated backpropagation through time to make training on long sequences more manageable, while retaining hidden states that can carry information from earlier observations.

The reported comparison finds that the RNN outperforms a vector autoregression baseline on a held-out period covering hundreds of stocks. The article also compares stock-specific models with a shared model trained across stocks. It reports that the shared model generally has better accuracy, less overfitting, lower computational cost, and more stable performance across stocks and sectors, including on unseen stocks. These findings come from the cited study and should be read as evidence for that dataset and setup, not a guarantee of trading profitability. The document reports direction-classification results; it does not establish returns after transaction costs or explain a deployable trading rule.

Key ideas

• The model predicts the direction of the next price change from high-frequency order-book sequences.
• Filtering to price-change observations reduces the volume of data used.
• Truncated backpropagation through time makes recurrent model training on long sequences more practical.
• The reported tests favor the RNN over a vector autoregression baseline.
• A shared model across stocks is reported to outperform many stock-specific models in most cases.

Tags

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