Impact of LLMs News Sentiment Analysis on Stock Price Movement Prediction

Battle Count: 135

The paper is relevant to quantitative trading as it evaluates how news sentiment features can improve stock movement prediction. However, the practical impact is limited: classification improvements are marginal (within standard deviations of baseline), and the study covers only 5 large-cap stocks. The regression improvements for PatchTST and TimesNet are more substantial. The findings suggest sentiment features provide modest but architecture-dependent benefits, which is useful information for practitioners building sentiment-augmented trading models. The paper does not address execution, portfolio construction, or risk management aspects critical for live trading.

The pipeline involves multiple stages: LLM inference for sentiment, daily aggregation, ensemble stacking, and integration with four different time-series architectures. However, all components use well-known open-source models and libraries. The hardware requirements are modest (consumer GPU with 4GB VRAM). The main complexity lies in the data pipeline (news collection, timestamp alignment, aggregation) and the combinatorial nature of experiments (4 architectures × 7 sentiment models × 2 tasks × 5 stocks × 10 seeds).

The authors provide a complete codebase on GitHub (https://github.com/Walids35/capstone-stock-prediction.git), specify hardware (NVIDIA GeForce GTX 1650 Ti, 4GB VRAM, 16GB RAM), use standard open-source Python libraries, run 10 experiments with different random seeds, and use publicly available datasets (Yahoo Finance, AlphaVantage, SEntFiN 1.0). Temporal train/val/test splits (70/10/20) are clearly defined. However, some hyperparameters for stock prediction models are not fully detailed in the paper.