TradeInbox: LLM-Based Real-Time Personalized Financial News Notification System
–
Supervisor: Prof. Chow, Ka Ho
Lee Jong Seung (3035555547)
Kim Taehyun (3035741330)
Lee Changjin (3035435840)
Project Introduction
In today’s fast-paced financial markets, which operates in real-time and is highly sensitive to news and sentiments, tracking news in real-time is crucial. Investor often find it challenging to determine the relevance of news to their own portfolios and to track such information efficiently, even in real-time.
To address this issue, we propose developing an LLM-based real-time personalized financial news alert application.
This application offers a website which allows users to input their investment portfolio and receive real-time notifications about relevant news with a summary of the news and an impact to the user portfolio. This ensures that users receive only the most relevant updates, enabling them to focus on news that matters for their financial decisions. Also, the content summary will help users quickly grasp key insights and suggest them which articles worth reading in full.
By providing timely and relevant updates, this project will help individual investors make informed and quick decisions, empowering them to stay ahead in the financial markets with ease.
Project Milestones
Phase 1: Inception (1 Oct, 2024 – 17 Jan, 2025)
1 Oct – 31 Oct, 2024
- Complete system architecture design and review on a full-scale
- Test Refinitiv API and validate any alternative data source if required
- Test embedding search model
1 Nov – 30 Nov, 2024
- Complete UI Design with Figma
- Database Schema Design
- Backend implementation of user authentication and stock keyword input functionality
- Infrastructure setup with AWS
1 Dec – 31 Dec, 2024
- Backend Implementation and integration with stock API
- Implementation of the Keywords Generator LLM
- Implementation of the embedding generator model
- Vector database infrastructure set up
- Front-end baseline setup
1 Jan – 31 Jan, 2025
- Frontend development
- Backend development
- Implementation of polling agent and integration with the message queue
- Implementation of the embedding search model
Phase 2: Elaboration (18 Jan, 2025 – 20 Apr, 2025)
1 Feb – 28 Feb, 2025
- Frontend development
- Backend development
- Implementation of prompt engineering on LLM summarization on news articles
- Implementation of Document Analyzer
1 Mar – 31 Mar, 2025
- Implementation of the backend of the LLM chat function
- Test and experiment with the Stock Analyzer LLM (summary generation LLM & stock price impact analysis NLP model)
1 Apr – 20 Apr, 2025
- Implementation of the frontend of the LLM chat function
- Integration test of the overall system
- Continuous testing and enhancement of the language models of the system
Phase 3: Construction (21 Apr, 2025 -)
21 Apr – 30 Apr, 2025
- Preparation of final presentation and project exhibition
Methodology & Results
Our Solution & Methodology
Intelligent Filtering: Using a unique Keyword Generator LLM combined with semantic embedding search, we identify news relevant to your stocks. This captures both direct mentions and crucial context (like industry trends or competitor news) often missed by simple keyword searches (using a 0.4 similarity threshold for balanced results).
AI-Powered Analysis: We don’t just deliver links. TradeInbox utilizes multiple prompt-engineered LLMs to provide:
- Concise article summaries.
- Key metric extraction.
- Sentiment analysis (Positive, Negative, Neutral).
- Multi-level stock impact analysis (Beginner, Intermediate, Expert) to match your financial understanding.
Real-Time Delivery: Get timely notifications with summaries delivered directly via Discord webhook.
Key Achievements
- Fully Functional System: We’ve successfully implemented an end-to-end platform featuring a user dashboard, detailed analysis pages, and real-time Discord alerts.
- Proven Relevance: Our filtering approach effectively identifies both directly and indirectly relevant news, validated through experiments showing meaningful similarity scores (0.4-0.6+) while excluding unrelated articles.
- Positive User Feedback: Initial testing showed strong user satisfaction (>70%), particularly praising the system’s promptness, personalization, educational value, and ease of comprehension.
The Impact
TradeInbox empowers retail investors by significantly reducing information overload, addressing information asymmetry, and enhancing financial literacy through clear, tailored, and actionable insights delivered in real-time.