Stock Price Prediction Method Based on Optimized LSTM and Quantified News Sentiment

Abstract

Predicting stock prices is a challenging task influenced by various factors, including historical trends and market sentiment. Traditional time series models, such as ARIMA, and conventional machine learning models, like Support Vector Machines (SVM) and K-Nearest Neighbors (KNN), often struggle to capture the nonlinear characteristics of the stock market, resulting in low forecasting accuracy [1]. This paper introduces a hybrid model that combines Long Short-Term Memory (LSTM) networks optimized with Genetic Algorithms (GA) and text sentiment analysis using pre-trained language models based on BERT and the cnsenti package. This approach effectively captures features from multimodal data, leading to improved predictions for Nvidia's stock price. To enhance the model's ability to learn from stock splits, we apply a price adjustment technique for continuity. The LSTM model, based on historical data, predicts stock prices, while the sentiment analysis model extracts insights from financial news. We utilize a weighted fusion approach to combine these predictions, providing a more comprehensive outlook on stock price movements. The final evaluation includes multiple performance metrics, such as Mean Squared Error (MSE), R-squared, Mean Absolute Error (MAE) and compares our model with general LSTM. Experimental results show that our hybrid method outperforms traditional LSTM models in terms accuracy and precision, offering a more robust framework for stock price forecasting.