Large

Modelsin

Finance:From

Innovationto

Real-World

Impact演講人

：劉煒清2017

2021Anti-Money

LaunderingWe

haveexpanded

our

industrial

partnershipintotheareas

of

Regtech,Anomaly

Detection,and

Fraud

Detection.We

haveappliedAItechniquesto

real-world

anti-money

laundering

scenario

andachieved

high

performance.

9

years

ofdeepcollaborationwithfinancial

industry

partners.Intelligent

InvestmentTogetherwithour

partners,we

launch

the

exploration

ofapplyingAItechniquesto

quantitative

investment. 9yearsofdeepcollaborationwithfinancial

industry

partners.

Researchtopicsfrom

realtasks.

Research

results

into

realproducts.2017201820192021Key

ResearchChallenges

ofAIfor

FinanceWe

havesummarized

andaddressed

several

key

research

challenges

inthefieldofAIfor

Finance.

Wealso

keep

updatingourtechniques

inpracticewithour

latest

research

findings.Anti-MoneyLaunderingWe

haveexpanded

our

industrial

partnershipintotheareas

of

Regtech,Anomaly

Detection,and

Fraud

Detection.We

haveappliedAItechniquesto

real-world

anti-money

laundering

scenario

andachieved

high

performance.TechniquesappliedOur

partners

startto

implement

ourresearch

andtechniques

intheir

real-world

productionswithgreatsuccess

and

reached

the

highest

levelof

performancethus

far.Intelligent

InvestmentTogetherwithour

partners,we

launch

the

exploration

ofapplyingAItechniquesto

quantitative

investment. 9yearsofdeepcollaborationwithfinancial

industry

partners.

Researchtopicsfrom

realtasks.

Research

results

into

realproducts.》sequentia

Decision

Making.

“universal

Trading

for

orderExecution

withoraclepolicyDistillation

”,

AAAI2021..“Towards

ApplicableReinforcementLearning:Improving

the

Generalizationand

sampleEficiency

withpolicyEnsemble

”,IJcAI

2022..“LearningMulti-AgentIntention

-Aware

communication

for

optimalMulti-orderExecutioninFinance

”,KDD2023.

“BpQp:

ADiferentiable

convex

optimizationFramework

forEficientEnd-to-EndLearning”,NeurIps2024.

Time-variant

pattern.“DDG-DA:DataDistributionGeneration

forpredictable

conceptDrift

Adaptation

”,

AAAI

2022..“LearningMultiple

stock

Tradingpatterns

with

TemporalRouting

Adaptor

andoptimal

Transport

”,KDD2021..

“Multi-Granularity

Residual

Learning

with

conidence

Estimation

for

Time

series

prediction

”,

www

2022.

comp

excorre

ation.“DeepRiskModel:

ADeepLearningsolution

forMiningLatentRiskFactors

toImprove

covarianceMatrixEstimation

”,IcAIF

2021.“Temporally

correlated

Task

scheduling

for

sequenceLearning

”,IcML2021..“Removingcamoulage

andRevealingcollusion:LeveragingGang-crimepatterninFraudsterDetection”,KDD2023.

Exp

ainabi

ity/lnterpretabiity.“LearningDiferential

operators

forInterpretable

Time

seriesModeling

”,KDD2022..“MeasuringModel

complexity

ofNeuralNetworks

withcurve

ActivationFunctions

”,KDD2020》Heterogenous

and

Hierarchica

Data.

“stock

Trendprediction

withMulti

-granularityData:

A

contrastiveLearning

Approach

with

AdaptiveFusion

”,cIKM2021..“REsT:RelationalEvent-drivenstock

TrendForecasting

”,

www

2021..“Digger-Guider:High-FrequencyFactorExtraction

for

stock

Trendprediction”,

TKDE2024..“MG-TsD:Multi-granularity

time

seriesdifusionmodels

with

guidedlearningprocess”,IcLR2024. 9yearsofdeepcollaborationwithfinancial

industry

partners.

Researchtopicsfrom

realtasks.

Research

results

into

realproducts.》》》sequentiaDecision

Making.“universal

Trading

for

orderExecution

withoraclepolicyDistillation

”,

AAAI2021..“Towards

ApplicableReinforcementLearning:Improving

the

Generalizationand

sampleEficiency

withpolicyEnsemble

”,IJcAI

2022..“LearningMulti-AgentIntention

-Aware

communication

for

optimalMulti-orderExecutioninFinance

”,KDD2023.“BpQp:

ADiferentiable

convexoptimizationFramework

forEficientEnd-to-EndLearning”,NeurIps2024.Time-variant

pattern.“DDG-DA:DataDistributionGeneration

forpredictable

conceptDrift

Adaptation

”,

AAAI

2022..“LearningMultiple

stock

Tradingpatterns

with

TemporalRouting

Adaptor

andoptimal

Transport

”,KDD2021..

“Multi-GranularityResidualLearning

withconidenceEstimation

for

Time

seriesprediction

”,

www

2022

.Heterogenous

and

Hierarchica

Data.“stock

Trendprediction

withMulti

-granularityData:A

contrastiveLearning

Approach

with

AdaptiveFusion

”,cIKM2021..“REsT:RelationalEvent-drivenstock

TrendForecasting

”,

www

2021

..“Digger-Guider:High-FrequencyFactorExtraction

for

stock

Trendprediction

”,TKDE2024..“MG-TsD:Multi-granularity

time

seriesdifusionmodels

withguidedlearningprocess”,IcLR2024.

comp

excorre

ation.“DeepRiskModel:

ADeepLearningsolution

forMiningLatentRiskFactors

toImprove

covarianceMatrixEstimation

”,IcAIF

2021.“Temporally

correlated

Taskscheduling

for

sequenceLearning

”,IcML2021..“Removingcamoulage

andRevealingcollusion:LeveragingGang-crimepatterninFraudsterDetection”,KDD2023.

Exp

ainabi

ity/lnterpretabiity.“LearningDiferentialoperators

forInterpretable

Time

seriesModeling

”,KDD2022..“MeasuringModelcomplexity

ofNeuralNetworks

withcurve

ActivationFunctions

”,KDD2020ResearchProduct 9yearsofdeepcollaborationwithfinancial

industry

partners.

Researchtopicsfrom

realtasks.

Research

results

into

realproducts.

Popularopensourcequant

investmentframeworkQlib.2017

2018

2019

2020

2021Key

ResearchChallenges

ofAIfor

FinanceWe

havesummarized

andaddressed

several

key

research

challenges

inthefieldofAIfor

Finance.

Wealso

keep

updatingourtechniques

inpracticewithour

latest

research

findings.Anti-MoneyLaunderingWe

haveexpanded

our

industrial

partnershipintotheareas

of

Regtech,Anomaly

Detection,and

Fraud

Detection.We

haveappliedAItechniquesto

real-world

anti-money

laundering

scenario

andachieved

high

performance.ReleaseofQlibWe

haveopen-sourced

our

researchtoolset

andplatform,Qlib,onGitHub.

It

supports

multiple

learning

paradigms

andcoverstheentire

process

ofquantitative

investment.

It

has

received

morethan

28kstars

sofar.TechniquesappliedOur

partners

startto

implement

ourresearch

andtechniques

intheir

real-world

productionswithgreatsuccess

and

reached

the

highest

levelof

performancethus

far.Intelligent

InvestmentTogetherwithour

partners,we

launch

the

exploration

ofapplyingAItechniquesto

quantitative

investment.Qlib

isanAI-orientedQuant

investment

platformthat

aimsto

useAItechtoempowerQuant

Research,

from

exploring

ideasto

implementing

productions.

Qlibsupportsdiverse

ML

modeling

paradigms,

includingsupervisedlearning(2020)Design

Goal:

Bridgingthegap

between

researchand

product

inquant

investment/microsoft/qlibQuant

ResearchAutomation

Powered

byQuant

Investment

FrameworkAdvantagesEasilyfusing

multi

-modalitydata

sourcesConcernsand

Limitations1.

Notdeterministic

result2.

Hard

to

be

evaluated3.

Hardtofurther

improve4.

Hard

to

be

regulatedAdvantagesHigher

intelligencewithin

financialdomainConcernsand

LimitationsHardto

handle

non

natural

languagefinancial

dataOurVisionFinanceAI2.0:

deterministicagentswithself-evolutionin

a

market-in-the-loopsystem.Code-OrientedAgentic

AutomationDomain-NativeFoundation

ModelsFinancialDomainFoundationModelOurApproachFinancialDomainAgento

Foundation

models

godomain-native.

Trainonorder-level

market

data

(notjust

text)tocapture

microstructure

andshow

scaling

behavior

realistic,controllable

generation

(LMM).o

Evaluationgoessimulation-centricMove

fromoffline

benchmarksto

market-in-the

-loop

stress

tests

and

what-ifs

policy,

risk,

andcompliance

measured

beforecapital

is

atrisk

(MarS).o

Agentic

R&D

goes

deterministic

LLMagents

generate

code,

run

backtests,and

auto-iteratewithobjective

metricspowerful,

repeatable,

reviewable

decisions

(Qlib

+

R&D-Agent).Code-basedSolution#3

#2

TrainedModel#1創(chuàng)R&D-AgentModelTraining324:2o

Foundation

models

godomain-native.

Trainonorder-level

market

data

(notjust

text)tocapture

microstructure

andshow

scaling

behavior

realistic,controllable

generation

(LMM).o

Evaluationgoessimulation-centricMove

fromoffline

benchmarksto

market-in-the

-loop

stress

tests

and

what-ifs

policy,

risk,

andcompliance

measured

beforecapital

is

atrisk

(MarS).Code-basedSolutionQuant

ResearchAutomation

poweredby

andRR&D-Agent#3

#2

TrainedModel#1創(chuàng)R&D-AgentModelTraining3242Quant

ResearchAutomationpowered

byQlibandR&D-Agent劉煒清微軟亞洲研究院IncollaborationwithXuYang,XiaoYang,Shikai

Fang,YugeZhang,ZehuaWang,YelongShen,WeizhuChen,Jiang

Biano

Buildingmain-stream

MLmodelsonspeciallydesignedtask

and

dataset.o

Continuousexpert-led

refinementthroughtrial,exploration,andadaptation.SolutionIntelligenceR&D

Agentcodifies

humanexperts'

R-D

Co

Evolution

LOOPSDevelopmentAgentImplementation

&

validationGenerateshigh-qualityideas

&experimentplansProducesrobust,production-ready

codeResearchAgentIdeation

&

hypothesis

craftingExternal

EnvironmentDevelopmentAgentResearchAgentFeedback

on

Trainingand

Executionmmmm

In-depthReason

of

R

mm

D

DesignImplementedSolutionFeedback

on

Quality?Typical

Quan

t

R&Dpipelineandopportunitiesof

Automation?Through

52

rounds

of

automated

evolution

over

18

hours,

a

comprehensively

enhancedintelligentquant

investmentsolution

has

beendeveloped,outperformingacrossallfour

key

metrics.?

Thegraydashed

lines

representthetopsolutionson

Qlib.?The

Automated

Evolvingof

Intelligent

QuantInvestmentsolutionmetricsachievedby

previouslyopen-sourcedat3Code-basedSolutionLarge

MarketModel

(

)

&

Its

Universal

Financial

MarketSimulation

Engine

(

)Quant

ResearchAutomation

powered

by

andRR&D-Agent#3

#2

#1TrainedModel創(chuàng)R&D-AgentModelTraining242Large

Market

Model(LMM)&

ItsUniversal

Financial

MarketSimulationIncollaborationwithJunjie

Li,Yang

Liu,ChangXu,Shikai

Fang,

LewenWangandJiang

BianEngine

(MarS)劉煒清微軟亞洲研究院Financial

MarketML

Modelfrom

MSRAo

Building

ML

modelsonmarketindicators,Indicators

of

the

marketInvestor/

Regulator?

Challengefortraditional

ML

methods?

Opportunityfor

Large

Foundation

Model★orderFlow

Market

Perspectivein

Macroscopic

Individual

Perspectivein

MicroscopicTokenizationofan

Order-BatchTokenizationofa

Single

Order…Transformer-basedAuto-regressiveTrainingonTwo-KindsofToken

SeriesLarge

Market

ModelIndividual

Perspective

in

MicroscopicMarket

Perspective

in

MacroscopicOrder-levelHistorical

Market

DataOrder-batch

ModelOrder

ModelTrainingLMMNew

ParadigmTraditionTraditional

Methods?Need

redesign

for

new

targets.NewApproach?Uses

recent

real

data

as

input

in

LMM.?Generatesfutureorderflows

and

corresponding

markettrajectories.?Derivesany

indicator’s

point/distributionalforecastingacrossdiversehorizons

from

multiple

future

markettrajectories.Enhanced

Performance?Outperformstraditionalalgorithms

in

predictions.?Evidenceof

LMM’s

powerful

modellingcapabilities.Accuracy0.70.60.50.41-min2-min

3-min

4-min

5-min3-class

Classification

of

PricePrediction

LMM

(1.02B)DeepLOB

(5different

models)Multiple

Rounds

ofGenerationIt’sallaboutspeed,speed,speed!~15xspeed-upofresponsetime

makes

thepapergainsto

real-world

promise.worldpromise*Accuracydropssignificantlywithfewerthan

16rollouts.ControlLMMInteractMarS:ControllableandInteractable

Financial

Market

SimulationPowered

by

LMMMarSDeployTrainVague

DescriptionofTargetScenarioUser-InjectedOrdersDigitalTwinExisting

orders

Generated

ordersincluding

user

submitted

orders

including

user

submitted

ordersO

O”Shaping

theFutureBased

onRealizedRealities”minutetcancelt-2t-1BidAsKcancelBidAsKBidtcancelAsK

o

vocabERefined

Logits”ReflectingImmediatechannelscancelBidAsK

×廠