Aligning Multilingual News for Stock Return Prediction

Battle Count: 128

Highly relevant to quantitative trading. The paper directly addresses stock return prediction using news text, constructs long-short trading strategies, and reports Sharpe ratios (up to 4.36 annualized for aligned Japanese embeddings). The multilingual alignment approach is novel for financial applications and could improve signal extraction from global news flows. However, the low correlation with realized returns (~0.02-0.03) and idealized Sharpe ratios suggest practical implementation challenges. The method is most applicable to cross-border equity strategies and multi-language news-driven trading systems.

Moderate to high complexity. Requires: (1) multilingual sentence embedding with LaBSE, (2) optimal transport computation via Sinkhorn algorithm with entropic regularization, (3) bidirectional alignment with thresholding, (4) embedding aggregation by alignment status, (5) Ridge regression with rolling window cross-validation, (6) long-short portfolio construction. The sparse-plus-low-rank optimization for OT adds computational sophistication. However, each component uses well-established libraries (spaCy, LaBSE, POT library for OT, sklearn for Ridge).

Bloomberg news data is provided under a non-disclosure agreement, making exact reproduction impossible. Stock return data from Compustat Global via WRDS is accessible. The methodology (LaBSE embeddings, Sinkhorn algorithm, Ridge regression) is well-described, but the specific hyperparameters (cosine threshold 0.6, top 5% selection) and preprocessing details would need to be replicated. No code repository is mentioned.