Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction

Battle Count: 182

Highly relevant to quantitative trading. The paper directly addresses stock return prediction and portfolio construction (long-only and long-short) using quantitative factors and newsflow. It provides practical insights on fusion methods, mixture models, and LLM fine-tuning effects across different market universes. The backtesting methodology (monthly rebalancing, decile-based portfolios) aligns with industry practice. Results show meaningful alpha generation (e.g., 32.43% annualized return for Fusion Combination in NA universe vs 12.37% universe return).

Moderate complexity. The fusion learning methods (combination, summation, attention) are relatively straightforward with dense layers. The mixture model with decoupled training adds complexity through the two-stage optimization (independent training + distribution matching via KL divergence). LLM integration with LoRA fine-tuning requires GPU resources (2x A100). The theoretical analysis (Propositions 1 and 2) is rigorous but the implementation uses standard PyTorch/HuggingFace components. Key complexity lies in the decoupled training scheme and ensuring proper gradient flow through LLM representations.

The paper provides detailed implementation descriptions including architecture specifics, hyperparameters (batch size 32, learning rate 1e-4, dropout 0.3, weight decay 1e-4, LoRA rank 4, max context length 4k, 10 epochs, 2x A100 GPUs), and uses publicly available DeBERTa. However, the financial news data is from a commercial vendor and not publicly available, and no code repository is mentioned. The factor data (~200 factors) and news categories are described but not fully specified.