Integration of LSTM Networks in Random Forest Algorithms for Stock Market Trading Predictions

Battle Count: 194

Directly addresses algorithmic stock trading with a practical intraweek system. Proposes a novel hybrid architecture combining sequential (LSTM) and static (Random Forest) data modalities. Includes simulation results comparing against major indices (S&P 500, NASDAQ, EuroStoxx 50). The feature selection approach (filtering by LSTM AUC) and the demonstration that LSTM predictions become the top feature when quality is high are practically relevant. However, the modest absolute AUC improvements and short validation period limit immediate deployability.

Requires training 482 individual LSTM networks (20-35 minutes each, totaling several weeks), followed by Random Forest training with augmented features. Uses standard Python libraries (Scikit-learn, Keras/TensorFlow). Greedy Search optimization for LSTM hyperparameters adds complexity. The hybrid architecture is conceptually straightforward but computationally intensive at scale. Hardware requirements are modest (consumer laptop sufficient for individual models) but the overall pipeline requires significant time investment.

The authors provide a GitHub repository with data, computer codes, and the list of 482 companies. However, the fundamental microeconomic data was provided by an anonymous multinational financial data company, limiting full reproducibility. Hyperparameter configurations are detailed, and standard libraries (Scikit-learn, Keras/TensorFlow) are used. The LSTM architecture uses a generalized configuration rather than per-asset optimization, which simplifies reproduction but may not capture optimal per-asset settings.