Cross-Sectional Heterogeneity in LSTM Networks for Financial Time Series

Battle Count: 151

Highly relevant to quantitative trading. The paper directly addresses daily directional stock prediction for S&P 500 constituents, constructs long-short portfolios, evaluates transaction costs, and provides risk metrics (VaR, CVaR, drawdown). The sector embedding approach provides a parsimonious way to incorporate cross-sectional heterogeneity. The decomposition into industry momentum and short-term reversal connects to well-known trading factors. However, the modest and regime-dependent nature of the signal, along with post-2008 performance deterioration, tempers practical applicability.

Moderate complexity. The base LSTM architecture is standard, and the sector embedding is a simple concatenation with the hidden state before the final linear layer (adding only 26 parameters for 2D embedding). The rolling-window training protocol with 27 study periods requires careful implementation. Regularization techniques (label smoothing, dropout, gradient clipping, weight decay, early stopping) are standard. The main complexity lies in data preprocessing: constructing a survivorship-bias-free S&P 500 universe, handling covariate gaps and outliers, and implementing the proper train/validation/test splits without data leakage. The interpretability analysis (latent space visualization, sector contribution decomposition) adds additional analytical complexity.

The paper states that all models together with their fitted parameters are available on GitHub. The methodology follows Fischer and Krauss (2018) with clearly described modifications. Data sources are specified (Refinitiv for S&P 500 constituents, Kenneth French Data Library for market returns, Bloomberg/Thomson Reuters for covariates). Hyperparameters are detailed (60-day lookback, 25 hidden units, 3 layers, Adam optimizer, 27 rolling study periods). However, the exact GitHub URL is not provided in the extract.