The paper introduces LLMFactor, a novel framework for explainable stock movement prediction that leverages Large Language Models (LLMs) to extract profitable factors from financial news. The core innovation is Sequential Knowledge-Guided Prompting (SKGP), a three-stage prompting strategy: first, it uses a fill-in-the-blank technique to elicit background knowledge (e.g., company relationships) from LLMs; second, it prompts LLMs to identify key factors that may affect stock prices from news articles; third, it integrates these factors, relations, and textualized historical price movements to predict whether the stock will rise or fall. This approach addresses limitations of prior methods that rely on keyphrases or sentiment analysis, which lack direct correlation with stock prices or human readability. Factors provide superior explainability by directly linking news content to price drivers.

Extensive experiments on four benchmark datasets (StockNet, CMIN-US, CMIN-CN, EDT) from US and Chinese markets demonstrate that LLMFactor consistently outperforms state-of-the-art methods, including keyphrase-based, sentiment-based, and time-series models. For instance, on StockNet, LLMFactor achieves 66.32% accuracy and 0.238 MCC, surpassing the previous best (CMIN) by 2.9% in MCC. Ablation studies confirm that each component (price data, factors, relations) contributes positively. The framework also provides interpretable explanations by listing the top factors influencing each prediction, enhancing trust and usability for financial analysts. The work highlights the potential of LLMs for financial time-series forecasting and factor analysis, with implications for trading strategy development and risk management.