Post

LLMQuant x HKUDS x FICC: AI 金融研究系统合作地图

LLMQuant x HKUDS x FICC: AI 金融研究系统合作地图

这是一篇专门整理:

1
LLMQuant x HKUDS x FICC

三者如何合作的 learning note。

前面我们已经分别做过:

1
2
3
4
5
6
7
8
LLMQuant:
  finance / quant domain layer

HKUDS:
  general AI infra / RAG / agent / evaluation layer

FICC:
  real financial market scenario

现在要把三者合起来。

一句话总结:

1
LLMQuant 提供金融量化研究语境,HKUDS 提供 AI infra / RAG / agent / evaluation 方法,FICC 提供真实金融市场任务场景。

更系统一点:

1
2
3
4
5
LLMQuant x HKUDS x FICC
  = finance domain knowledge
  + AI research infrastructure
  + real market workflow
  + public-safe research artifacts

公开边界:

1
2
This is educational material and research infrastructure thinking.
It is not investment advice, trading advice, or an actionable alpha note.

中文边界:

1
2
这是公开学习笔记和研究系统设计,不是投资建议。
不包含内部数据、客户信息、未脱敏策略、实盘观点或可交易 alpha。

总体判断

LLMQuant、HKUDS、FICC 不是三条孤立线。

它们可以组成一个很完整的 AI Finance Research OS:

1
2
3
4
5
6
7
8
FICC:
  真实金融研究任务

LLMQuant:
  金融语义、量化任务、策略研究、数据接口、agent workflow

HKUDS:
  RAG、Graph、Agent、Harness、Evaluation、Workspace、Artifact

合在一起:

1
2
3
4
5
6
FICC event
  -> LLMQuant finance understanding
  -> HKUDS RAG / agent / harness infra
  -> quant hypothesis
  -> validation and bias diagnosis
  -> public-safe research artifact

我们要做的不是简单“套项目”。 我们要做的是:

1
用 FICC 作为真实场景,把 LLMQuant 和 HKUDS 的能力重新组织成一个可运行、可复盘、可展示的研究系统。

三者角色分工

方向角色代表内容在系统里的作用
FICC场景层FI / FX / Commodities / Daily Brief / Event-to-Signal提供真实金融研究问题
LLMQuant金融领域层QuantMind / data-mcp / Skills / Magents / awesome-trading-agents提供 finance / quant workflow
HKUDSAI 基础设施层LightRAG / RAG-Anything / AutoAgent / OpenHarness / DeepResearch-Eval提供 RAG、agent、evaluation、harness

可以压成:

1
2
3
FICC asks the question.
LLMQuant provides finance-native structure.
HKUDS provides AI-native infrastructure.

中文:

1
2
3
FICC 出题。
LLMQuant 给金融结构。
HKUDS 给 AI 基础设施。

Layer 1: Data Layer

FICC 需要的数据包括:

1
2
3
4
5
6
7
8
9
10
rates
yield curve
credit spreads
FX spot / forward / swap
cross-currency basis
commodity futures
inventory reports
macro calendar
central bank text
news and events

LLMQuant 里最直接可合作的是:

1
data-mcp

它对应 FICC 的数据接口层:

1
2
3
4
FICC market data
  -> MCP-style data access
  -> agent-readable schema
  -> daily brief / event store / validation engine

HKUDS 对应的是工具化和接口化:

1
2
3
4
CLI-Anything
AnyTool
AutoAgent
AgentSpace

这些项目启发我们:

1
2
不要让 agent 靠自然语言猜数据。
要让 agent 通过明确工具接口拿数据。

FICC 合作形态:

1
2
3
4
5
6
7
FICC Data MCP
  - rates endpoint
  - FX endpoint
  - commodities endpoint
  - macro calendar endpoint
  - event store endpoint
  - forecast ledger endpoint

Layer 2: Knowledge Layer

FICC 最大的问题不是没有信息。 是信息太散:

1
2
3
4
5
6
7
央行 statement
macro releases
EIA inventory report
WASDE
research reports
market commentary
historical daily briefs

LLMQuant 里最重要的是:

1
2
QuantMind
finance knowledge layer

它适合做:

1
2
3
4
5
FICC concept cards
FICC event cards
FICC report cards
FICC hypothesis cards
FICC strategy logic cards

HKUDS 里最适合合作的是:

1
2
3
4
LightRAG
RAG-Anything
MiniRAG
VideoRAG

对应能力:

1
2
3
4
5
6
7
8
9
10
11
LightRAG:
  fast graph-aware retrieval

RAG-Anything:
  multimodal document understanding: PDF, table, chart, formula

MiniRAG:
  lightweight retrieval memory

VideoRAG:
  视频、访谈、课程、讲座进入知识系统

FICC 合作形态:

1
2
3
4
5
FICC Knowledge RAG
  = QuantMind finance schema
  + LightRAG graph retrieval
  + RAG-Anything document ingestion
  + MiniRAG lightweight memory

用途:

1
2
3
4
5
Daily Brief evidence retrieval
central bank stance retrieval
commodity report parsing
historical analogies
counter-evidence search

Layer 3: Graph Layer

FICC 是天然图结构。

关系包括:

1
2
3
4
central bank -> policy rate -> yield curve -> FX
oil price -> inflation -> rates -> FX -> credit
inventory shock -> futures curve -> commodity FX
USD liquidity -> EM FX -> credit spreads

LLMQuant 的合作点:

1
2
QuantMind semantic knowledge graph
finance-domain entity and relation schema

HKUDS 的合作点:

1
2
3
4
GraphAgent
OpenGraph
GraphGPT
HiGPT

它们启发:

1
2
不要只做 text RAG。
要做 FICC cross-asset graph。

FICC Graph 可以包含节点:

1
2
3
4
5
6
7
8
9
10
11
12
Asset
Currency
Country
CentralBank
MacroVariable
Commodity
Curve
Event
Claim
Hypothesis
Evidence
ValidationReport

关系:

1
2
3
4
5
6
7
8
9
Event -> affects -> Asset
CentralBank -> sets -> PolicyRate
PolicyRate -> affects -> YieldCurve
YieldCurve -> affects -> FXForward
CommodityShock -> affects -> Inflation
Inflation -> affects -> Rates
Evidence -> supports -> Hypothesis
CounterEvidence -> challenges -> Hypothesis
ValidationReport -> evaluates -> Hypothesis

合作形态:

1
2
3
4
FICC Graph RAG
  = QuantMind finance semantics
  + HKUDS graph intelligence
  + FICC cross-asset causality

Layer 4: Agent Layer

FICC Research OS 需要多个 agent。

LLMQuant 合作点:

1
2
3
4
Skills
Magents
awesome-trading-agents
strategy and risk workflows

HKUDS 合作点:

1
2
3
4
5
6
AutoAgent
AgentSpace
OpenSpace
nanobot
FastAgent
ClawTeam / ClawWork

这些项目可以启发我们设计:

1
2
3
4
5
6
7
8
9
10
11
Rates Agent
FX Agent
Commodities Agent
Macro Agent
Daily Brief Agent
Event Agent
Hypothesis Agent
Validation Agent
Skeptic Agent
Redaction Agent
Ledger Agent

关键不是 agent 多。 关键是 agent 必须有 harness:

1
2
3
4
5
6
fixed task
fixed input
fixed output schema
evidence required
counter-evidence required
human review required

FICC 合作形态:

1
2
3
4
FICC Multi-Agent Research Workflow
  = LLMQuant finance task routing
  + HKUDS agent workspace
  + FICC domain agents

Layer 5: Harness Layer

这是 FICC008 直接对应的层。

LLMQuant 的 finance agent 如果要用于 FICC,必须有:

1
2
3
4
5
6
data boundary
tool boundary
memory boundary
output schema
bias diagnosis
public-safety review

HKUDS 里最相关的是:

1
2
OpenHarness
DeepResearch-Eval

OpenHarness 对应:

1
2
3
4
5
task contract
tool permission
schema validation
review gate
audit log

DeepResearch-Eval 对应:

1
2
3
4
5
research quality evaluation
evidence grounding
answer faithfulness
counter-evidence coverage
human edit distance

FICC 合作形态:

1
2
3
4
FICC Agent Harness
  = OpenHarness-style control plane
  + LLMQuant finance workflow
  + FICC public-safe policy

这个方向非常重要。

因为金融 agent 最大问题不是“能不能生成内容”。 而是:

1
能不能在边界内生成可审计的研究内容。

Layer 6: Quant Workflow Layer

FICC 最后要进入 quant research。

LLMQuant 直接相关:

1
2
3
4
Magents
awesome-trading-agents
strategy workflows
factor / strategy hypothesis direction

HKUDS 相关:

1
2
3
4
5
6
Vibe-Trading
AI-Trader
FutureShow
AI-Researcher
Auto-Deep-Research
DeepResearch-Eval

FICC 对应:

1
2
3
4
Event-to-Signal Workflow
Forecast Ledger
Validation Report
Bias Diagnosis

合作形态:

1
2
3
4
5
FICC Event-to-Signal Engine
  = LLMQuant quant strategy workflow
  + HKUDS AI-Researcher loop
  + HKUDS DeepResearch-Eval
  + FICC event store

标准流程:

1
2
3
4
5
6
7
8
FICC event
  -> hypothesis
  -> feature
  -> outcome
  -> event study
  -> bias diagnosis
  -> research verdict
  -> forecast ledger

公开边界:

1
signal candidate != trading signal

公开展示的是:

1
2
3
4
research hypothesis
validation framework
bias checklist
synthetic demo

不是:

1
2
3
live alpha
position sizing
execution instruction

Layer 7: Evaluation Layer

没有 evaluation,Research OS 只是内容生成器。

LLMQuant 可以提供:

1
2
3
4
5
finance-specific evaluation:
  data validity
  backtest health
  strategy risk check
  factor robustness

HKUDS 可以提供:

1
2
3
DeepResearch-Eval
OpenHarness
research agent evaluation

FICC 需要的 evaluation:

1
2
3
4
5
6
7
8
9
Daily Brief Quality
Evidence Grounding
Counter-Evidence Coverage
Cross-Asset Consistency
Timestamp Correctness
Bias Diagnosis Quality
Forecast Calibration
Human Review Pass Rate
Public-Safety Compliance

合作形态:

1
2
3
4
FICC Research Evaluation Suite
  = DeepResearch-Eval style metrics
  + OpenHarness review gate
  + LLMQuant finance robustness checks

这能把 FICC 系统从 demo 提升成真正 research infra。

Layer 8: Artifact Layer

研究要沉淀成可展示资产。

LLMQuant 可以输出:

1
2
3
4
5
finance research notes
strategy specs
factor cards
risk reports
quant workflow docs

HKUDS 可以参考:

1
2
3
4
Paper2Slides
Litewrite
DeepResearch reports
agent product workflow

FICC 可以输出:

1
2
3
4
5
6
7
Daily FICC Brief
FICC Event Card
FICC Signal Candidate Card
FICC Bias Report
FICC Forecast Ledger
FICC AI Research OS Demo
FICC Agent Harness Spec

合作形态:

1
2
3
4
5
FICC Public Artifact Layer
  = public-safe research outputs
  + website learning posts
  + GitHub project docs
  + private research ledger separation

这直接服务我们的 Credit OS。

总比对表

FICC 需求LLMQuant 侧HKUDS 侧合作结果
数据接口data-mcpCLI-Anything / AnyToolFICC Data MCP
金融知识QuantMindLightRAG / RAG-AnythingFICC Knowledge RAG
图谱关系QuantMind semantic KGGraphAgent / OpenGraph / GraphGPTFICC Graph RAG
Agent workflowSkills / MagentsAutoAgent / AgentSpace / OpenSpaceFICC Multi-Agent Workflow
Harness 控制finance workflow rulesOpenHarnessFICC Agent Harness
研究自动化quant R&D directionAI-Researcher / Auto-Deep-ResearchFICC R&D Agent
交易研究Magents / awesome-trading-agentsVibe-Trading / AI-TraderFICC Event-to-Signal
预测复盘research ledger directionFutureShowFICC Forecast Ledger
评估backtest / risk checkDeepResearch-EvalFICC Research Evaluation
公开表达strategy notesPaper2Slides / LitewriteFICC Public Artifact Layer

优先合作路线

如果按我们当前目标排序,优先级是:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Priority 1:
  QuantMind + LightRAG + RAG-Anything
  -> FICC Knowledge RAG

Priority 2:
  data-mcp + AnyTool / CLI-Anything
  -> FICC Data MCP

Priority 3:
  OpenHarness + DeepResearch-Eval
  -> FICC Agent Harness / Evaluation

Priority 4:
  AI-Researcher + Magents + Vibe-Trading
  -> FICC Event-to-Signal Workflow

Priority 5:
  FutureShow
  -> Forecast Ledger / Prediction Evaluation

一句话:

1
先做知识和 harness,再做 research workflow,再做 quant validation。

不要一上来就冲实盘。

先把:

1
2
3
4
5
knowledge
evidence
harness
evaluation
ledger

打牢。

一个具体合作 Demo

可以做一个公开安全 demo:

1
FICC Daily Research OS Demo

输入:

1
2
3
4
5
sample market snapshot
sample macro calendar
sample central bank statement
sample EIA-style inventory note
sample FICC learning notes

系统:

1
2
3
4
5
6
LLMQuant data-mcp style input
QuantMind-style finance knowledge card
HKUDS LightRAG-style retrieval
HKUDS GraphAgent-style cross-asset graph
HKUDS OpenHarness-style control policy
HKUDS DeepResearch-Eval-style evaluation

输出:

1
2
3
4
5
6
7
8
Daily FICC Brief
Event Cards
Signal Candidate Cards
Evidence Pack
Counter-Evidence Pack
Bias Diagnostic Report
Forecast Ledger Update
Public-safe Website Post

这个 demo 非常适合展示:

1
AI agent + FICC domain + quant research + public-safe engineering

对外叙事

对外可以这样讲:

1
2
I am building a public-safe FICC AI Research OS.
It combines LLMQuant-style finance research workflow, HKUDS-style RAG / Graph / Agent / Evaluation infrastructure, and real FICC market tasks such as daily brief generation, event-to-signal workflow, forecast ledger, and human-reviewed research artifacts.

中文:

1
2
我在构建一个公开安全的 FICC AI Research OS。
它把 LLMQuant 的金融量化研究工作流、HKUDS 的 RAG / Graph / Agent / Evaluation 基础设施,以及 FICC 的真实金融研究场景结合起来,形成 daily brief、event-to-signal、forecast ledger 和 human-reviewed research artifact 的闭环。

面试表达

如果被问:

1
LLMQuant、HKUDS 和 FICC 怎么结合?

可以回答:

1
我会把 FICC 看成真实金融场景,把 LLMQuant 看成 finance-native research workflow,把 HKUDS 看成 general AI research infrastructure。具体来说,LLMQuant 的 QuantMind 和 data-mcp 可以提供金融知识和数据接口;HKUDS 的 LightRAG、RAG-Anything、GraphAgent 可以提供检索和图谱能力;OpenHarness 和 DeepResearch-Eval 可以提供 agent 控制和评估;AI-Researcher、Vibe-Trading、AI-Trader 可以启发 event-to-signal 和 quant validation workflow。最终合成一个 FICC AI Research OS:从 public-safe data ingestion,到 daily brief,到 event-to-signal,到 validation、bias diagnosis、forecast ledger 和 human review。

如果被问:

1
这个系统和普通金融 RAG 有什么区别?

可以回答:

1
普通金融 RAG 主要回答问题或总结文档。这个系统不止做问答,而是把 RAG 作为 research memory,把 Graph RAG 作为 cross-asset relationship layer,把 agent workflow 作为研究执行层,把 quant validation 作为验证层,把 harness 和 human review 作为控制层,把 forecast ledger 作为复盘层。所以它是一个研究操作系统,而不是单点 RAG 应用。

当前结论

LLMQuant、HKUDS、FICC 的合作可以压成:

1
2
3
4
5
6
7
8
LLMQuant:
  finance domain and quant workflow

HKUDS:
  AI infra, RAG, graph, agent, harness, evaluation

FICC:
  real financial research scenario

最终合成:

1
2
3
4
5
FICC AI Research OS
  = FICC scenario
  + LLMQuant finance workflow
  + HKUDS AI infrastructure
  + public-safe research artifacts

一句话收束:

1
LLMQuant gives the finance brain, HKUDS gives the AI infrastructure, and FICC gives the real market training ground.

中文:

1
LLMQuant 给金融大脑,HKUDS 给 AI 基础设施,FICC 给真实市场训练场。
This post is licensed under CC BY 4.0 by the author.