FANAL slashes financial news processing costs to 0.001% of GPT-4 while maintaining superior accuracy
FANAL is a specialized BERT-based framework for real-time financial event detection that categorizes news into 12 distinct categories. It uses silver-labeled data through XGBoost and employs ORPO fine-tuning, creating ORBERT for superior class-wise probability calibration, outperforming GPT-4 and other LLMs at fraction of the cost.
-----
https://arxiv.org/abs/2412.03527
🔍 Original Problem:
→ Financial markets generate massive news streams that overwhelm traditional analysis methods
→ Existing LLMs are computationally expensive and impractical for real-time financial news processing
→ Current solutions lack granular categorization needed for precise decision-making
-----
💡 Solution in this Paper:
→ FANAL processes financial news through a BERT model fine-tuned with ORPO technique
→ It uses XGBoost for creating silver-labeled training data from just 1,200 manually labeled samples
→ The framework introduces ORBERT, optimizing class-wise probability calibration
→ Entity Relevance Module enhances news interpretation for specific financial signals
-----
🎯 Key Insights:
→ Silver labeling with minimal manual data can achieve competitive performance
→ ORPO fine-tuning significantly improves class balance handling
→ Smaller, specialized models can outperform larger LLMs in specific domains
-----
📊 Results:
→ Processes 10,000 articles in 0.013 hours vs GPT-4's 1.579 hours
→ Costs $0.0017 for 10,000 articles vs GPT-4's $204.87
→ Achieves 96%+ accuracy in categories like Bankruptcy and Strategic Alliances
Share this post