Sentiment Analysis of Financial News Articles

Classifying Articles into Positive and Negative Sentiments

---

Objective

• Classify financial news articles based on sentiment.
• Distinguish between positive and negative sentiments.
• Compare neural network and boosting models for accurate classification.

---

Dataset Overview

• Dataset: Collection of financial news articles.
• Columns:
  • Sentiment: Label indicating sentiment (positive or negative).
  • Sentence: The textual content of the news article.
• Total Samples: Approximately 2000.
• Preprocessing: Cleaning, tokenization, and encoding of text data.
• Imbalanced Data: Addressed to ensure balanced class distribution.

---

Methodology

1. Data Cleaning and Preprocessing

   • Removal of HTML tags, special characters, and stopwords.
   • Tokenization and normalization of text.
   • Encoding textual data for model compatibility.

2. Handling Class Imbalance via Oversampling

   • Upsampling the minority class to balance the dataset.
   • Ensuring equal representation of positive and negative sentiments.

3. Feature Extraction using Tokenization and TF-IDF

   • Converting text data into numerical features.
   • Utilizing Tokenization and Term Frequency-Inverse Document Frequency (TF-IDF) for feature representation.

4. Model Building

   • Convolutional Neural Network (CNN)
   • XGBoost Classifier
   • Random Forest Classifier

5. Hyperparameter Tuning

   • CNN: Tuned dropout rates using Keras-Tuner's RandomSearch.
   • XGBoost: Employed RandomizedSearchCV for optimizing learning rate, max-depth, n_estimators, etc.
   • Random Forest: Optimized parameters like n_estimators, max_depth, etc.
   • Objective: Enhance model performance and prevent overfitting.

6. Evaluation Metrics

   • Accuracy
   • Precision
   • Recall
   • F1 Score
   • AUC (Area Under the Curve)

---

CNN Architecture

• Embedding Layer: Converts words into dense vectors of fixed size.
• Convolutional Layer: Captures local features in the text.
• MaxPooling Layer: Reduces dimensionality and captures the most significant features.
• Dropout Layers: Prevent overfitting by randomly dropping neurons during training.
• Dense Layers: Learns complex patterns in the data.
• Global Max Pooling: Aggregates feature maps into a single vector.
• Output Layer: Sigmoid activation for binary classification.

---

Hyperparameter Tuning

• CNN Model:

  • Tuned dropout rates using Keras-Tuner's RandomSearch to find optimal values that balance model complexity and generalization.

• XGBoost:

  • Utilized RandomizedSearchCV to optimize hyperparameters such as learning rate, max-depth, number of estimators, and minimum child weight.

• Random Forest:

  • Optimized parameters like the number of estimators and maximum depth to enhance performance and reduce overfitting.

• Objective:

  • Enhance overall model performance.
  • Prevent overfitting through appropriate hyperparameter settings.

---

Model Performance Comparison

Model	Train Accuracy	Test Accuracy	Precision	Recall	F1 Score	AUC
Random Forest	0.72	0.67	0.67	1.00	0.80	0.88
XGBoost	0.99	0.86	0.88	0.94	0.91	0.80
CNN	0.99	0.85	0.83	0.96	0.89	0.92

Evaluation Metrics for Models

Detailed evaluation metrics including Accuracy, Precision, Recall, F1 Score, and AUC for each model are provided in the Model Performance Comparison table above.

---

Novelties of the Approach

• Combined Deep Learning with Traditional ML Models:

  • Integrated Convolutional Neural Networks (CNN) with boosting models like XGBoost and ensemble methods like Random Forest.

• Advanced Hyperparameter Tuning:

  • Implemented sophisticated hyperparameter tuning techniques using Keras-Tuner and RandomizedSearchCV to achieve optimal model performance.

• Comprehensive Evaluation Metrics:

  • Applied a wide range of evaluation metrics to thoroughly assess model robustness and effectiveness.

• Effective Handling of Class Imbalance:

  • Addressed class imbalance through oversampling techniques to improve classification accuracy and model fairness.

---

Conclusion

• Performance Achievement:

  • Achieved over 80% accuracy in sentiment classification of financial news articles.

• Model Comparison:

  • Performances of CNN and XGBoost are comparable, with XGBoost performing marginally better.

• Hyperparameter Tuning Impact:

  • Effective hyperparameter tuning significantly enhanced model performance. Further tuning could enable CNN to outperform XGBoost.

---