.png){kind=icon}
.png){kind=icon}
Return
Secure Multibranch Deep Learning Framework for Effective Stock Sentiment Analysis
.png){kind=icon}
PRE
AI
DOI：10.1109/tcss.2026.3691026.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The efficient market hypothesis (EMH) posits that stock prices reflect all available information, including financial news. Still, existing stock sentiment analysis models overlook three fundamental drivers of market reaction, namely, the significance, intensity, and duration of news. This study introduces a multibranch deep learning framework that explicitly incorporates these dimensions to generate market-aligned sentiment predictions. Central to the framework is a manual sentiment voting (MSV) mechanism that derives sentiment labels from stock return series, thereby quantitatively capturing directional, magnitude-based, and persistence-adjusted market responses to news events. Numerical and textual information are modeled through dedicated pipelines that integrate technical indicators, Hurst exponent analysis, and discrete wavelet transform for numerical data, and FinBERT embeddings for financial text, followed by 1-D-convolutional neural network (CNN) and long short-term memory (LSTM)-based feature extraction. Robustness and privacy are ensured through local differential privacy, character anonymization, and adversarial training. Experiments on two extensive real-world data show that the framework achieves 98.07% accuracy, while an extensive ablation study confirms the critical contribution of each component to overall performance.
Keywords:
Adversarial learning
financial news
FinBERT
convolutional neural network (CNN)
long short-term memory (LSTM)
Stock sentiment
Journal
IF:
4.9
Papers:
577
Citations:
6.8K