Financial Sentiment Analysis

A Natural Language Processing (NLP) project that classifies financial news and statements into sentiment categories using both traditional machine learning and deep learning approaches.

Overview

This project uses the Financial PhraseBank dataset to predict sentiment labels:

• Positive
• Neutral
• Negative

The objective is to compare different NLP pipelines and evaluate their effectiveness on financial text classification.

Dataset

Dataset: Financial PhraseBank

The dataset contains financial news sentences annotated with sentiment labels by human experts.

Features

• Text preprocessing and cleaning
• Exploratory Data Analysis (EDA)
• TF-IDF feature extraction
• Machine Learning baselines
• LSTM-based deep learning model
• Performance comparison across models
• Training and validation loss visualization

Models Evaluated

Traditional Machine Learning

• Naive Bayes
• Logistic Regression(TF-IDF and WordToVec)
• LSTM (WordToVec)
• FinBERT

Deep Learning

• LSTM Classifier (PyTorch)

Project Structure

.
├── finance-phrasebank/
├── models/
├── utils/
├── bulk_test/
├── eda_scikit_learn_improved.ipynb
├── bulk_evaluate.py
├── requirements.txt
└── README.md

Technologies Used

• Python
• Pandas
• NumPy
• Scikit-learn
• PyTorch
• Matplotlib
• Jupyter Notebook

Training

Example:

train_infer_lstm(
    X_train_lstm,
    y_train_lstm,
    epoch=9
)

Evaluation Metrics

Models are evaluated using:

• Accuracy
• Precision
• Recall
• F1 Score
• Confusion Matrix

Results

Model	Accuracy
Logistic Regression	XX%
SVM	XX%
Random Forest	XX%
LSTM	XX%

Replace with your actual results.

Key Learnings

• Compared classical NLP pipelines with neural network approaches.
• Explored the impact of sequence-based models on sentiment classification.
• Implemented batching, embedding layers, and LSTM architectures in PyTorch.
• Evaluated model performance using multiple classification metrics.