Explainable Prediction of Economic Time Series Using IMFs and Neural Networks

Battle Count: 186

The paper is moderately relevant to quantitative trading. It provides insights into which temporal components (long-term trends vs. high-frequency oscillations) drive neural network predictions of stock prices, which could inform feature engineering and model interpretation in trading systems. The explainability aspect is valuable for understanding model behavior before deploying in production. However, the paper does not address trading strategies, transaction costs, portfolio construction, or risk-adjusted returns. The predictive accuracy of the LSTM model is notably poor, and the MLP's near-perfect R² may indicate overfitting to the specific datasets. The work is more foundational/academic than directly applicable to trading systems.

The methodology involves several components: EMD decomposition (available in libraries like PyEMD), neural network training (MLP and LSTM via Keras/TensorFlow), and DeepSHAP attribution (via shap library). Each component is individually well-supported by existing Python libraries. The main complexity lies in correctly configuring the EMD decomposition, ensuring proper data splitting for IMFs, and interpreting DeepSHAP results. The architectures are relatively simple (single hidden layer MLP, single LSTM layer), making implementation straightforward for practitioners familiar with deep learning frameworks.

The paper describes model architectures (MLP with 64 hidden neurons, LSTM with 10 memory units), training parameters (Adam optimizer, learning rate 0.001, batch size 64, early stopping), and data splitting (75/25). However, no code repository is provided, no specific random seeds are mentioned, and the EMD implementation details are not fully specified. The datasets (NVIDIA and Apple stock prices) are publicly available, which aids reproducibility.