Generalized Stock Price Prediction for Multiple Stocks Combined with News Fusion

Battle Count: 74

The paper directly addresses stock price prediction, a core task in quantitative trading. The generalized multi-stock model and news fusion approach are relevant for signal generation in trading strategies. However, the paper focuses on price level prediction (MAE/MSE) rather than directional accuracy, trading signals, or portfolio-level performance. No backtesting, transaction costs, or trading strategy evaluation is included. The 7.11% MAE reduction is meaningful but the practical trading edge is not demonstrated. The computational cost (2 days training) may limit real-time retraining frequency.

The pipeline involves multiple complex components: BERT/DeBERTa encoding of 200+ daily news articles, three attention-based pooling variants, bidirectional cross-attention fusion, two-layer GCN with CausalCNN, patch reprogramming into LLM embedding space, and fine-tuning of reprogramming layers while keeping the LLM frozen. Requires GPU infrastructure (RTX-3090, 24GB), familiarity with HuggingFace transformers, Time-LLM framework, and graph neural networks. The multi-stock training setup and news collection pipeline add further engineering complexity.

The paper provides detailed model architecture, training configurations (batch size, learning rate, epochs, early stopping), dataset descriptions, and evaluation metrics. However, no GitHub repository or code link is provided. The datasets (TW21 via twstock API, BigData23) are publicly accessible. Model weights (Llama3-TAIDE-LX-8B, gpt2-base-chinese, BERT, DeBERTa) are available on HuggingFace. The exact news collection pipeline and preprocessing steps are described but not fully reproducible without the authors' code.