LLM4Finance
May 2023 - Current
Developing an LLM-powered finance dashboard for personal
investment with additional financial engineering techniques.
Pipeline Outline
MLOps Pipeline for Sentiment Analysis
- PEFT is applied to LLM to predict the sentiment of crawled news headlines: positive, neutral, or negative.
- Continuous training is triggered in a HuggingFace space environment when the amount of newly labeled data surpasses a specified threshold.
- Continuous deployment then follows in a separate HuggingFace space environment.
- The trained model is quantized with llama.cpp framework for deployment in a free basic CPU space of HF.
- The HuggingFace model hub is used to save and load trained models.
LLM Application
- LLM agents are crafted for data processing processes data: auto-labels sentiment data and controls the quality of crawled news headlines.
- Error analysis revealed that data quality issues, such as punctuation and duplication, can severely impact model performance. LLM agents were crafted to address this.
- The LLM assistant is designed to enhance and streamline financial statement analysis, providing comprehensive support and insights.
- Experimental: RAG-based framework for enhanced detailed analysis.
- Experimental: Multiple LLM agents are crafted to simulate economic investment environment to respond to specific scenarios, each tailored to different investing personas.