QuantBrain-Agent: Building a Fully Automated A-Share Research System with Qwen2.5-72B That Delivers Reports at 9 AM Sharp Every Day

While Wall Street is still manually scrolling through K-line charts on Bloomberg Terminals, someone has already built an AI that delivers institutional-grade research reports to your inbox at 9 AM every morning. The open-source project we're discussing today — QuantBrain-Agent — is a fully automated end-to-end investment research system for China's A-share market, built on the Qwen2.5-72B-AWQ large language model. It directly tackles the three biggest pain points in equity research: slow turnaround, incomplete analysis logic, and difficulty quantifying qualitative data.

Project link: SR88888888/quantbrain-agent

---

The Three Pain Points of A-Share Research — How Bad Are They Really?

Anyone who's done A-share research knows the traditional workflow suffers from three chronic problems:

• Slow turnaround: By the time an analyst finishes writing a report, the stock has already hit its daily limit up. Markets move in milliseconds, but manually collecting data, organizing logic, and drafting reports takes far too long.
• Incomplete logic: Too many research reports dare to issue buy recommendations based on P/E alone, lacking multi-dimensional cross-validation and supply chain correlation analysis.
• Hard-to-quantify data: Market sentiments and qualitative judgments like "good news is already priced in" are extremely difficult to express and model with structured numbers.

QuantBrain-Agent has ambitious goals — it aims to automate the entire research pipeline, essentially replacing a small team comprising data analysts, industry researchers, and report writers with AI. If this system truly works as intended, brokerage research departments might be turning off their lights a few hours earlier.

---

Core Technical Architecture: A Greatest Hits Collection of 2024's Hottest AI Papers

This project's tech stack is impressively stacked, packaging nearly every cutting-edge AI technique from the past two years into one comprehensive system. Let's break it down piece by piece.

Qwen2.5-72B-AWQ: Stuffing an Elephant into a Fridge — And It Can Still Dance

The system's core inference engine uses Qwen2.5-72B from Alibaba's Tongyi Qianwen team — a 72-billion parameter large language model with strong language understanding and generation capabilities.

But 72 billion parameters means enormous VRAM requirements, making direct deployment prohibitively expensive. The project employs AWQ (Activation-aware Weight Quantization) to solve this problem. AWQ's core insight is clever: not all weights have equal impact on model output. By observing activation values during inference, it identifies which weight channels are most critical, gives those important weights special protection, and aggressively compresses the rest to 4-bit precision. The result: dramatically reduced model size and VRAM requirements with minimal accuracy loss.

Think of it as stuffing an elephant into a fridge — except miraculously, the elephant can still dance.

LangGraph Multi-Agent Collaboration: Drawing a Workflow Diagram for AI Employees

No single AI, however powerful, can handle complex investment research tasks alone. QuantBrain-Agent uses the LangGraph framework to orchestrate collaboration among multiple AI agents.

LangGraph, developed by the LangChain team, is a multi-agent workflow framework that models the entire AI application as a directed graph — nodes represent different processing steps or AI agents, edges represent flow direction, with support for conditional branching and loops.

Essentially, it's drawing a clear workflow diagram for a team of AI employees: who's responsible for finding data, who handles analysis, who writes the report, who reviews for errors — clear division of labor, each agent in its designated role. Compared to traditional linear chain calls, this graph structure enables more complex collaboration patterns, such as loop-back flows where a failed review triggers revisions.

LoRA Vertical Fine-tuning: Giving a General LLM a Crash Course in A-Shares

General-purpose LLMs know a little about everything, but they're not specialized enough for the vertical domain of financial research. The project uses LoRA (Low-Rank Adaptation) for domain-specific fine-tuning.

LoRA's principle is elegant: traditional fine-tuning requires updating all model parameters, which is extremely costly. Research has shown, however, that when a model adapts to new tasks, the weight changes are actually "low-rank" — they can be approximated by the product of two very small matrices. So LoRA freezes the original model weights and inserts two small matrices alongside each layer, reducing trainable parameters by a factor of 1,000 or even 10,000.

This is like giving Qwen2.5-72B a crash course in A-shares, transforming it from "knows a little about everything" to "knows finance exceptionally well" — more familiar with financial terminology, better at analytical reasoning, and more aligned with A-share market characteristics.

AWQ quantization + LoRA fine-tuning is a common and smart deployment strategy for large models: first compress with AWQ to a deployable size, then adapt with LoRA for domain expertise, balancing deployment cost with domain specialization.

MCP Tool Chain: AI's Universal Remote Control

MCP (Model Context Protocol) is an open standard protocol proposed by Anthropic in November 2024. Think of it as the "USB interface" for AI.

Through MCP, large models can invoke various external tools using a unified, standardized approach — market data APIs, news scraping tools, financial data queries, and more — without needing custom adapter code for each tool. In the investment research context, this means the system can fetch the latest stock prices, financial statements, and news in real-time, rather than relying solely on "stale knowledge" from training.

CoT + ReAct: Think While Searching, Search Then Think Again

For the reasoning architecture, the project combines CoT (Chain of Thought) and ReAct (Reasoning + Acting).

• CoT makes the model show its work step by step rather than jumping to conclusions, like an analyst writing out every logical step from data to conclusion in a research report.
• ReAct enables the model to alternate between "thinking" and "acting" during reasoning — when it realizes it needs certain data, it goes and fetches it; when new information comes in, it continues analyzing, looping until reaching a conclusion.

Combined, the model no longer works in isolation making gut-feeling conclusions. Instead, it operates like a real researcher: think while searching, search then think again, think then verify.

The design philosophy behind this entire architecture is: if one AI isn't enough, bring in a whole team; still not enough? Teach them to find their own data and self-reflect.

---

Complete Pipeline: From Data Collection to Report Delivery

QuantBrain-Agent isn't just a tech demo — it builds a complete investment research production line:

Real-time Market and Sentiment Data Collection

The system uses MCP tool chains to capture market data and sentiment information in real-time, solving the "information bubble" problem in traditional research. Whether it's unusual stock price movements, policy changes, or trending discussions on social media, everything can be incorporated into the analysis.

Industry Chain Knowledge Graph Multi-dimensional Reasoning

This is one of the most interesting parts of the system. Knowledge graphs organize information using "entity-relationship-entity" structures, making them particularly suited for modeling supply chain relationships.

For example: if lithium ore prices rise, the system can quickly trace through the knowledge graph to identify all affected battery manufacturers, automakers, and even further downstream to charging station companies and energy storage sectors. This kind of multi-dimensional correlation analysis is extremely difficult to achieve with traditional single-threaded research.

Reflexion Self-Review: AI Critiquing Its Own Work

The most elegant design is the Reflexion mechanism. After generating a report, the system doesn't output it directly — instead, it performs self-evaluation and reflection, checking whether the logic is complete, the data is accurate, and the conclusions are sound. Issues are automatically corrected, and this process can iterate multiple times.

Reflexion's workflow is: execute task → self-evaluate → identify shortcomings → store in memory → re-execute. In plain terms, the AI writes the report, then proofreads it, finds its own mistakes, and revises — showing more self-awareness than some interns.

Daily 9 AM Output with Automatic Feishu Push

Finally, the system produces institutional-grade professional research reports at 9:00 AM every day and automatically pushes them to Feishu (Lark). This setup is practically a dream for fund managers — getting a multi-dimensional analysis report before market open. For sell-side analysts, though, it might be less welcome news.

The "scheduled output + Feishu push" detail also indicates this isn't a toy project stuck in the proof-of-concept stage, but a system designed for actual production environments.

---

The Thinking Behind the Technology Choices

Looking back at the overall tech stack, several design choices are worth noting:

1. CoT + ReAct + Reflexion Trinity: CoT handles reasoning depth, ReAct handles interaction with the external environment for real-time information, and Reflexion handles quality control. Together they form a complete "reasoning-acting-reflecting" closed loop — the current best practice for mainstream Agent architectures.

2. AWQ + LoRA Combo: First quantize to lower the deployment barrier, then fine-tune to boost domain capability — a pragmatic strategy that balances cost and effectiveness.

3. LangGraph Over Simple Chain Calls: Investment research tasks inherently require multi-step, multi-role collaboration. Graph-structured workflows are far more flexible than linear pipelines, especially in scenarios requiring "review-revise" loops.

---

Final Thoughts

The name "quantbrain" — "quantitative brain" — perfectly captures this project's positioning: an AI agent with quantitative analysis capabilities. It integrates large language models, multi-agent systems, fine-tuning, quantization, and knowledge graphs into the vertical domain of financial research, embodying the current trend of AI applications moving from general capabilities toward deep industry-specific deployment.

Of course, whether AI-generated research reports can truly reach "institutional-grade" quality still requires continuous validation in real-world use. But at least in terms of architectural completeness and engineering maturity, this project offers a highly convincing reference implementation.

In the AI era, what's truly terrifying isn't that machines are smarter than you — it's that they're more diligent AND better at self-reflection. The AI submits its homework at 9 AM sharp every day, while you're still debating whether to add sugar to your coffee.

Project link: https://github.com/SR88888888/quantbrain-agent