Financial News Anlysis for stock prediciton with Machine Learning approachs

📌 Project Overview

Objective

Develop an NLP-powered sentiment analysis engine to predict stock movements using financial news data to help traders in their daily decisions.

Methodology

1. Sentiment Analysis ✅​
   --> Analyze the sentiments and perspectives within the texts to discern the global sentiment of a document​

2. Topic Modeling​ ✅ --> Identify and categorize key themes in the texts to obtain a comprehensive understanding of the topics.​

3. Aspect Based Sentiment​ ✅ ​--> Analyze specific aspects in the text to understand sentiment nuances or specific perspectives .​ ​

4. Historical Sentiment Records​ (Next steps) --> Explore temporal trends in sentiment analysis to understand the evolution of opinions over time ​ ​

mermaid
graph TD
    A[Raw Financial News] --> B(Sentiment Analysis)
    B --> C(Topic Modeling)
    C --> D(Aspect-Based Analysis)
    D --> E(Temporal Trend Analysis)
    E --> F[Stock Movement Prediction]

🔍 Core Components

1. Sentiment Analysis Models

Model	Architecture	Epochs	Classes	Accuracy
DistilRoberta	Transformer	10	2	100% (overfit)
FinancialBERT	BERT-base	10	2	99% (overfit)
FinancialLarge	RoBERTa-large	5	3	92%

What I chose to finetune:

Two-Class Sentiment Analysis:​​

• Data classified into Negative or Positive sentiments​​
• Fine-tuning with a 4-layer architecture​​
• Training conducted over 10 epochs​

Three-Class Sentiment Analysis:​​

• Data categorized into Negative, Neutral, and Positive sentiments​​
• Fine-tuning with a 4-layer architecture​​
• Model trained for 5 epochs​​

The results weren't good with the fine tunned models, it decreased some performance, especially the 3 class predictions. Let's moove on with the topic detection.

2. Topic Modeling Approach

I tested 3 LDA approachs. Linear discriminant analysis (LDA) is an approach used in supervised machine learning to solve multi-class classification problems. LDA separates multiple classes with multiple features through data dimensionality reduction.

Model	Pros	Cons
LDA 1	Robust, complete	Buggy code
LDA 2 (BART)	StackExchange fine-tuned	Very slow
LDA 3 (Gensim)	Clear, interactive visu	Alpha balancing tricky

Gensim visualisations with alpha manipulation :

Hybrid Solution choosen:

• LDA for probabilistic topic assignment
• BERT for semantic embedding
• Weighted concatenation (α = 0.4)

Key Parameters:

• 10 main financial topics
• 3 subtopics per category

3. Aspect-Based Analysis (ABSA)

Targeted goal: Be able to obtain a balanced analysis for a relevant measure of a sentiment associated to a speciic company or theme in a sentence where the opinion are mixed.

Aspect-Based Sentiment Analysis: Unlike general sentiment analysis, ABSA focuses on extracting sentiments associated with specific aspectsmentioned in a text.

Example Output:

{
  "text": "Bank A reported 15% revenue growth but increased debt",
  "aspects": [
    {"term": "revenue", "sentiment": "positive", "score": 0.82},
    {"term": "debt", "sentiment": "negative", "score": -0.76}
  ]
}

After identifying aspects, the model determines the sentiment (positive, negative, neutral) for each aspect in each document.

• Version that identify the aspect by extracting nouns:

• Version after LDA topics injected

Results & Deliverables

Achievements

• Good detection results with a big increase with LDA topic injection
• Aspect-level sentiment extraction

Challenges

• PhraseBank dataset limitation
• Dependency chain between modules
• Alpha parameter sensitivity in LDA-BERT hybrid

Next Steps

1. More data integration
2. Web scraper for historical sentiment
3. Unified prediction engine
4. Real-time API deployment

Dataset Example

{
  "sentence": "Orion Corp reported Q3 earnings fall due to R&D costs",
  "label": "negative",
  "topics": ["pharma", "earnings"],
  "aspects": {
    "earnings": -0.68,
    "R&D": 0.12
  }
}

Technical Stack

• NLP Models: HuggingFace Transformers, Gensim, SpaCy
• Visualization: PyLDAvis, Matplotlib
• Infrastructure: Python 3.9, PyTorch 2.0

Aurélien Pouxviel, Data Scientist.