From News to Trends: A Financial Time Series Forecasting Framework with LLM-Driven News Sentiment Analysis and Selective State Spaces

Stock price prediction is inherently challenging due to market volatility and the influence of external factors. Traditional forecasting methods primarily rely on historical price data, limiting their ability to capture market sentiment embedded in financial news. To address this limitation, we propose Senti-MambaMoE, a novel model that integrates historical stock prices with sentiment information extracted from financial news. Specifically, we fine-tune a DeepSeek-based large language model (LLM) for financial sentiment classification and incorporate the extracted sentiment information into our predictive framework. At the core of our approach is MambaMoE, which leverages the efficiency of state space models (SSMs) to model long-range dependencies while maintaining linear computational complexity, making it well-suited for financial time series forecasting. Additionally, the MoE mechanism improves the model’s ability to capture diverse market behaviors by dynamically selecting specialized experts based on stock data patterns. Experimental results demonstrate that Senti-MambaMoE outperforms LSTM-based models by 23.7% and Transformer-based models by 6.3%, highlighting its superior performance in short-term stock prediction.