Introduction
Financial news serves as a critical source of real-time market insights, influencing investor sentiment and stock prices. This study explores the relationship between V-A (Valence-Arousal) sentiment divergence in financial news and sector-based stock price movements, proposing a predictive model that enhances accuracy by incorporating multidimensional emotional analysis.
Literature Review
Financial News Impact on Stock Markets
Research indicates that sentiment extracted from financial news significantly affects stock volatility. Studies by Xu Wei [6] and Zhao Cheng et al. [7] demonstrate improved prediction accuracy using sentiment indicators alongside traditional stock metrics.
Sentiment Analysis Limitations
Traditional methods relying on average sentiment intensity often overlook emotional diversity, leading to information loss. Siganos et al. [5] highlight the importance of sentiment divergence in capturing nuanced market reactions.
V-A Sentiment Model
The Valence-Arousal model quantifies emotions along two continuous dimensions:
- Valence (V): Positivity/negativity (range: 0β9).
- Arousal (A): Calmness/excitement (range: 0β9).
This approach outperforms discrete models by distinguishing overlapping emotions [16].
Methodology
1. V-A Sentiment Calculation
A CNN-LSTM hybrid model analyzes news text to compute V-A scores:
- CNN Layer: Extracts local n-gram features.
- LSTM Layer: Captures long-range semantic dependencies.
- Linear Activation: Outputs continuous V-A values (Equation 1β2).
π Explore advanced sentiment analysis techniques
2. Sentiment Divergence Metric
Normalized V-A scores classify news as positive (V > 0) or negative (V < 0). Divergence is calculated using:
- Entropy-based measures (Equation 6).
- Daily sector-level aggregation of sentiment variability.
3. SVR Prediction Model
Inputs:
- Lagged sentiment divergence.
- Historical stock prices.
Optimization: GridSearchCV tunes hyperparameters (C, gamma) for minimal MAE/MSE.
Experimental Results
Data Sources
- News: 4,820 articles (Sohu Finance, 2016).
- Stocks: RESSET Financial Database.
- Preprocessing: Jieba segmentation + custom stopwords.
Key Findings
- Full-text sentiment divergence ([EDS]) achieved lowest error (MAE: 0.3829).
- Model outperformed title-only analysis ([TEDS]) by 22%.
- Sector C (Manufacturing) predictions shown below:
| Model | MSE | MAE |
|---|---|---|
| EDS | 0.203 | 0.383 |
| ES [3] | 0.241 | 0.421 |
| TEDS | 0.261 | 0.440 |
π See real-world trading applications
Conclusion
This study demonstrates that V-A sentiment divergence significantly improves sector-level stock predictions. Future work could integrate semi-supervised learning and additional textual features (e.g., syntactic patterns) for enhanced robustness.
FAQs
Q1: Why use V-A over discrete sentiment models?
A1: Continuous V-A scales capture nuanced emotions (e.g., mixed positive-negative states), reducing information loss.
Q2: How is sentiment divergence calculated?
A2: Daily entropy of sector-normalized V-A scores, aggregated from news articles (Equation 6).
Q3: Can this model predict individual stocks?
A3: Currently optimized for sectors; adapting toδΈͺθ‘ requires finer-grained news tagging.
Q4: What are the limitations?
A4: Dependence on labeled training data; future work may leverage unsupervised techniques.