README_EN.md

April 3, 2026 · View on GitHub

AlphaTeam

Building a trustworthy, controllable, and traceable open architecture for quantitative investment research in capital markets A Trustworthy, Controllable, and Traceable Framework for Financial Intelligence.

📖 Background & Lineage

AlphaTeam is the industrial-grade evolution of our previous academic open-source project, AlphaFin.

AlphaFin (Predecessor): Focused on the construction of high-quality financial datasets, LLM fine-tuning, retrieval augmentation, and preliminary explorations into single-agent chain-of-logic reasoning. It was designed to address the foundational challenges of financial data timeliness and computational complexity.
AlphaTeam: Building upon the financial intuition and data foundation established by AlphaFin, this project comprehensively upgrades to an architecture of Multi-Agent Collaboration and Systematic Engineering Governance. It marks the definitive leap from "heuristic single-agent Q&A" to a "trustworthy, controllable, and evolvable quantitative research framework."

💡 Core Philosophy

In noisy capital markets, we do not pursue a single-model “oracle.” We pursue deterministic research processes.

AlphaTeam addresses key pain points of traditional AI research systems: black-box workflows, unclear accountability, and weak auditability. We shift the focus from “result-only outputs” to interpretable processes, constrained actions, and traceable states.

AlphaTeam Architecture and Core Philosophy

🚀 Key Features

Feature	Description
🛡️ Reliable	A causal hypothesis-chain reasoning pattern enforces strict evidence-to-conclusion mapping to reduce hallucinations at the logic level.
🎮 Controllable	Built-in adversarial audit and risk re-check mechanisms challenge every investment conclusion under rigorous logic constraints.
🔍 Traceable	Full-chain Activity Stream plus memory center allows audit reconstruction of decision basis, tool calls, and source data.
🧬 Evolving	Deeply adapted to A-share first constraints; supports multi-agent evolution through real-world feedback loops.

🎬 Demo Preview

The following screenshot shows AlphaTeam in action, including visual collaboration and research workspace:

AlphaTeam Demo

🛠️ Architecture

AlphaTeam is not a simple chat assistant. It is an industrial-grade substrate for investment research:

1. Multi-Agent Collaboration Layer

Research, intelligence, risk, and audit responsibilities are separated to simulate real financial research pipelines with professional specialization.

2. Tool Governance Center

API calls and data retrieval are governed with strict auditing to ensure compliant sources, transparent call chains, and reproducible research outputs.

3. Visual Workflow

Decision chains are rendered transparently so quantitative logic is white-box and each strategy detail is visible.

🆚 AlphaTeam vs Traditional Financial Frameworks

Dimension	Traditional Finance / Traditional AI Research Frameworks	AlphaTeam
Execution Paradigm	Monolithic model or fixed rule chain, limited flexibility	Multi-agent collaboration + orchestrator-driven workflow
Controllability (process & policy governance)	Fragmented control planes; hard to unify timeout, approval, and interruption policies	Workflow-level configuration, timeout governance, global stop, and role-boundary constraints
Traceability	Results may be visible but processes are often opaque	End-to-end traceability via Activity Stream + Trace + report archival
Compliance & Audit Evidence	No unified audit entry; fragmented evidence chain	Unified evidence chain via tool catalog, source review, and audit APIs
Memory Transparency	Context memory is often hidden	Visualized memory center via `memory_center`
Visualization Capability	Mostly text and static charts	3D workflow scene + real-time status panel
Evolution Capacity	High extension cost, weak role reuse	Modular roles/tools/skills for continuous evolution

Mission: AlphaTeam does not replace human investment judgment. We provide a scalable, verifiable, and financially grounded digital infrastructure for global research teams.

Role Taxonomy (Academic Naming)

Runtime ID	Runtime Name	Academic Role
`director`	Decision Director	Coordinator
`analyst`	Investment Analyst	Fundamental Researcher
`intel`	Market Intelligence Officer	Market Intelligence Researcher
`quant`	Quant Strategist	Quantitative Researcher
`risk`	Risk Officer	Risk Reviewer
`auditor`	Reflective Auditor	Audit Reviewer
`restructuring`	Asset Restructuring Specialist	Special Situations Researcher

Architecture Spec

Full specification: docs/ARCHITECTURE.md. Summary:

Orchestration Layer: orchestrator and portfolio_scheduler drive research and portfolio cycles.
Agent Runtime Layer: agent_registry manages model routing, runtime states, and stop control.
Tool and Data Layer: tool_registry constrains tool entry points and parameter schemas.
Memory and Governance Layer: memory system, activity bus, traces, and tool audits form a governance loop.
Interface Layer: /team provides unified interaction and visualization entry, including 3D collaboration scenes.

Core Modules

App Entry: AlphaFin/app.py
Team API and Routes: AlphaFin/ai_team/routes.py
Orchestrator: AlphaFin/ai_team/core/orchestrator.py
Portfolio Scheduler: AlphaFin/ai_team/core/portfolio_scheduler.py
Agent Registry: AlphaFin/ai_team/core/agent_registry.py
Tool Registry: AlphaFin/ai_team/core/tool_registry.py
Memory System: AlphaFin/ai_team/core/memory.py
Team Frontend: AlphaFin/ai_team/templates/ai_team.html
Team 3D Scene: AlphaFin/ai_team/static/js/three_scene.js
DB Build Script: AlphaFin/scripts/build_db.py

📊 Indicator and Strategy Coverage (Brief)

The current open-source release includes 25 runnable indicator/strategy modules across four groups: Capital Flow (7), Macro and Valuation (7), Market Structure (7), and Strategy Models (4).
Current Strategy Models include: Industry Rotation, Quality Stock Screening, Agent Collaborative Strategy, and Master Stock-Picking Strategies.
Indicators follow a unified registry and loading path: auto-discovered by indicator_registry, and served/rendered through /indicator/<indicator_id>.
The design goal is that each indicator is independently usable while remaining auditable, reviewable, and reproducible in AlphaTeam workflows.

Quick Start

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure environment variables (see .env.example), at least:

SECRET_KEY
TUSHARE_TOKEN
QWEN_API_KEY
MOONSHOT_API_KEY

Start service:

python3 AlphaFin/app.py

Access:

Team Workspace: http://127.0.0.1:5002/team

Data Bootstrap

Default database directory is data/db (override with ALPHAFIN_DB_ROOT).

# Quick mode (recommended, recent 1 year)
python3 -m AlphaFin.scripts.build_db --mode quick

# Full mode
python3 -m AlphaFin.scripts.build_db --mode full

# Full mode + financial indicators
python3 -m AlphaFin.scripts.build_db --mode full --include-fina

Current release prioritizes A-share research context:

China market clock semantics and trading-session constraints.
Controllable, rebuildable local data pipelines.
Risk review and audit are default pre-steps.

The architecture can be extended to other markets by adding market-specific data adapters, clock rules, and trading semantics.

Observability and Governance APIs

Capability	API
Activity Stream (SSE)	`/api/team/activity`
Runtime Trace	`/api/team/trace/runs`, `/api/team/trace/<run_id>`
Memory Center	`/api/team/memory_center`
Tool Catalog and Source Review	`/api/team/tools_audit/catalog`, `/api/team/tools_audit/source`, `/api/team/tools_audit/review`
Lifecycle Controls	`/api/team/module/start`, `/api/team/module/stop`, `/api/team/stop_all_work`
Timeout Governance	`/api/team/session/<session_id>/overtime`

Non-goals

This project does not guarantee investment returns or win rates.
This project is not an automated live-trading system.
Outputs from this project do not constitute investment advice.

Citation

If you use AlphaTeam in research or engineering, please cite as software:

@software{alphateam2026,
  title   = {AlphaTeam: Trustworthy, Controllable, and Traceable Quantitative Trading PlatForm},
  author  = {AlphaTeam Contributors},
  year    = {2026},
  url     = {https://github.com/jackyideal/AlphaTeam}
}