返回 Papers
AI 扩展计划 / Playbooks

AI Treasury / Liquidity / ALM Forecasting / Stress Evidence Playbook

本文是学习、作品集和架构设计材料, 不构成法律意见、监管解释、模型验证结论、流动性充足性结论、资本/财务确认、投资建议、融资执行建议、会计意见、审计意见或董事会治理意见。

595AI_TREASURY_LIQUIDITY_ALM_FORECASTING_STRESS_EVIDENCE_PLAYBOOK.md

AI Treasury / Liquidity / ALM Forecasting / Stress Evidence Architecture Playbook

适用对象: CBAP-level Financial Retail PM / Senior BA / Treasury Product Owner / ALM Architect / Risk Data Product Lead / AI Governance / Model Risk / Enterprise Architect / Board MI Lead。 目标: 把 treasury、liquidity forecasting、deposit beta/runoff、cash-flow ladder、ALM、IRRBB、FTP、portfolio scenario、stress testing、contingency funding、human committee、board MI 和 data lineage 设计成可落地、可运营、可审计的 AI 产品架构能力。 核心观点: Treasury AI 的交付物不是预测数字, 而是 forecast-to-action system: 数据血缘、模型证据、压力假设、解释、人工挑战、行动记录和董事会管理信息。


0. Boundary And Disclaimer

本文是学习、作品集和架构设计材料, 不构成法律意见、监管解释、模型验证结论、流动性充足性结论、资本/财务确认、投资建议、融资执行建议、会计意见、审计意见或董事会治理意见。

正式项目中的适用范围、监管口径、报告义务、模型治理要求、客户数据使用、披露和委员会权限由 Legal、Compliance、Regulatory Affairs、Model Risk、Treasury、Finance、Risk、Internal Audit 和管理层权责方确认。本文只提供产品与架构设计框架。

Product boundary:

  • AI 可以预测、解释、模拟、生成候选情景、提醒异常、起草 ALCO/board MI 初稿。
  • AI 不应独立批准或执行 funding action、asset sale、hedging、FTP policy change、customer pricing、contingency funding activation、监管报告结论或董事会风险偏好变更。
  • 任何影响资金、流动性、客户、财务报告、监管报告、风险偏好或重大经营决策的动作都必须进入 human committee、dual control、limit check 和 evidence process。

1. Executive Framing

弱项目通常这样定义:

Build an AI model to forecast deposits and liquidity.

成熟项目应该这样定义:

Build a governed treasury intelligence and action-evidence platform
that connects cash-flow forecasts, deposit behavior assumptions,
ALM/IRRBB scenarios, FTP incentives, stress testing, CFP actions,
committee decisions and board MI.

Executive one-liner:

This is not a predictive analytics dashboard. It is a controlled liquidity decision system with evidence.

1.1 Steering Committee Questions

  1. 哪些 treasury / ALM 决策会被 AI 输出影响: daily liquidity, ALCO, FTP, pricing, portfolio, CFP, board MI?
  2. 每个预测的 grain、horizon、cadence、owner、allowed action 和 prohibited use 是否清楚?
  3. Deposit beta/runoff 是按什么 segment、scenario、rate regime 和 relationship evidence 定义?
  4. Cash-flow ladder 是否区分 contractual、behavioral、operational、contingent 和 intraday liquidity?
  5. ALM / IRRBB 的 NII/EVE 结果如何使用 AI assumptions, 谁批准 assumption change?
  6. FTP 是否把 funding risk、liquidity premium 和 contingent liquidity cost 反馈到产品/业务线激励?
  7. 压力情景、人工覆盖和 committee challenge 是否能被重放?
  8. Board MI 的每个数字是否可以追溯到数据、模型、情景、阈值、行动和 owner?

2. Source Anchors

访问日期: 2026-06-30。

AnchorOfficial link本 playbook 使用方式
Federal Reserve SR 10-6, Interagency Policy Statement on Funding and Liquidity Risk Management用户给定路径: https://www.federalreserve.gov/supervisionreg/srletters/sr1006.htm ; Fed legacy anchor: https://www.federalreserve.gov/boarddocs/srletters/2010/sr1006.htm ; attachment PDF: https://www.federalreserve.gov/boarddocs/srletters/2010/sr1006a1.pdf用 liquidity risk governance、stress testing、cash-flow measurement、management reporting、liquid asset cushion、diversified funding、CFP 和 internal control 组织控制框架。正式引用前复核用户给定路径可访问性。
Federal Reserve SR 11-7, Guidance on Model Risk Managementhttps://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm用 model inventory、validation、ongoing monitoring、model use、effective challenge 和 model uncertainty 设计 forecast/stress model control。
OCC Comptroller's Handbook indexhttps://www.occ.treas.gov/publications-and-resources/publications/comptrollers-handbook/index-comptrollers-handbook.html作为正式项目中定位 OCC Treasury / Market Risk、Interest Rate Risk、Liquidity、Model Risk 相关 handbook 的入口。
OCC Interest Rate Risk revised booklet bulletinhttps://www.occ.treas.gov/news-issuances/bulletins/2020/bulletin-2020-26.html用 IRR、NII/EVE、NMD assumptions、stress testing、model risk、FTP 语言校准 ALM/IRRBB 设计。
Federal Reserve SR 16-3, Interagency Guidance on Funds Transfer Pricing Related to Funding and Contingent Liquidity Riskshttps://www.federalreserve.gov/supervisionreg/srletters/sr1603.pdf用 FTP 连接 funding risk、liquidity component、contingent liquidity risk、business-line incentives 和 product profitability。
FFIEC Management booklethttps://ithandbook.ffiec.gov/it-booklets/management.aspx用 IT governance、risk management、resources、performance、change、monitoring 和 assurance 设计 technology operating controls。
NIST AI RMFhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织 AI 风险、评估、监控、human oversight、incident 和 evidence lifecycle。
ISO/IEC 42001https://www.iso.org/standard/42001用 AI management system、policy、roles、operation、performance evaluation、internal audit 和 continual improvement 设计 enterprise operating model。
Federal Register final policy statementhttps://www.federalregister.gov/documents/2010/03/22/2010-6137/interagency-policy-statement-on-funding-and-liquidity-risk-management作为 SR 10-6 policy statement 的发布锚点, 支持 liquidity risk management、stress testing、management reporting 和 CFP 控制讨论。

Source-to-control pattern:

official source
  -> control objective
  -> product requirement
  -> technical control
  -> evidence artifact
  -> owner
  -> monitoring metric

3. Target Operating Model

Treasury AI 要作为 cross-functional control product 管理。

3.1 Operating Forums

ForumPrimary purposeAI product responsibility
Treasury daily managementcash positioning, funding needs, collateral movement within approved limitsdaily forecast, intraday alerts, limit status, action evidence
ALCOliquidity, ALM, IRRBB, FTP, balance-sheet strategyscenario pack, assumption changes, NII/EVE, FTP and portfolio impacts
Treasury Risk / Market Riskindependent challenge, limits, stress reviewchallenger outputs, limit breach evidence, issue tracking
Model Risk Committeemodel inventory, validation, ongoing monitoringmodel pack, validation evidence, monitoring, change approval
AI Governance / AIMS forumenterprise AI policy, controls, incident, performanceAI RMF / ISO 42001-aligned control evidence
Board Risk / Audit Committeeoversight, risk appetite, management action challengeboard MI data product and evidence lineage

3.2 RACI

ActivityTreasuryALMRiskModel RiskDataTechnologyLegal/ComplianceInternal Audit
Forecast contract approvalACCCCCCI
Deposit assumption ownerCACCCIII
Stress scenario libraryAAACCCCI
Model validationCCCACCII
Data lineage and qualityCCCCARII
Action executionACCIIRCI
Board MI sign-offAAACCCCI
Control testingIICCCCCA

Legend: A = accountable, R = responsible, C = consulted, I = informed.


4. Reference Architecture

1. Source and lineage layer
   core banking, deposits, loans, payments, cards, GL, treasury trades,
   securities, collateral, market data, pricing, CRM, legal entity, limits

2. As-of treasury data products
   deposit balance and flow facts
   contractual cash-flow schedules
   behavioral cash-flow features
   payment and settlement calendars
   rate curve and competitor rate products
   collateral and liquidity source inventory

3. AI / analytics layer
   baseline forecast
   deposit beta / runoff
   cash-flow forecast
   anomaly / early-warning
   scenario generation
   challenger models
   explanation services

4. Risk and finance engines
   liquidity ladder
   stress survival horizon
   ALM / IRRBB NII and EVE
   FTP allocation and simulation
   portfolio scenario and limit impact

5. Decision workbench
   Treasury dashboard
   ALCO scenario workbench
   action recommendation with constraints
   override and committee challenge
   dual-control action register

6. Evidence and MI plane
   forecast contracts
   model inventory and validation
   scenario versioning
   assumption register
   data lineage graph
   action log
   board MI metric contracts
   audit export

Architecture rule:

If a forecast can change a liquidity action, it must have a contract, owner, lineage,
model-risk status, scenario context, decision boundary and evidence trail.

5. Core Product Modules

5.1 Treasury Data Product Layer

ModuleWhat it providesEvidence
Deposit balance and flow martdaily/intraday balances, inflows/outflows, product, segment, legal entitysource mapping, cutoff, reconciliation, restatement log
Cash-flow schedule servicecontractual loan/security/funding/coupon/maturity flowssource contract, schedule version, exception handling
Payment and settlement calendarACH/wire/card/check/payroll/tax/holiday settlement timingcalendar version, rail owner, holiday logic
Rate and market data productFed/market curves, product rates, competitor rates, spreadssource, timestamp, curve construction
Collateral and liquidity source inventoryunencumbered securities, eligibility, haircut, operational availabilitycustody source, encumbrance, legal entity
Customer concentration graphcustomer group, relationship, top depositor exposure, concentrationentity resolution logic, group owner
Limit and risk appetite registryliquidity, IRRBB, concentration, model, data and action limitsthreshold owner, approval date, review cadence

5.2 Forecast And Scenario Layer

ModuleProduct behavior
Baseline model suitenaive/seasonal, business rules, statistical and ML baselines for each forecast contract
Deposit beta/runoff enginesegment and scenario-level behavioral assumptions with uncertainty
Cash-flow forecast enginecontractual + behavioral + operational + contingent forecasts
Early-warning monitoranomalous flow, concentration movement, rate sensitivity, competitor gap, event signals
Scenario librarybase, idiosyncratic, market-wide, combined, reverse stress, operational and intraday
Challenger frameworkindependent model or rules that challenge primary output
Explainability servicedriver bridge, segment contribution, scenario sensitivity, action implication

5.3 Decision And Evidence Layer

ModuleProduct behavior
ALCO workbenchcompare scenarios, limits, NII/EVE, liquidity ladder, FTP impacts and actions
Action registerproposed action, owner, approval forum, constraint check, execution status, closure evidence
Overlay workflowhuman adjustment with reason, scope, expiry, approver and impact
Committee challenge logchallenge question, response, decision, residual risk owner
Board MI generatormetric contracts, thresholds, status, drivers, actions and evidence links
Audit evidence exportselected sample can reconstruct source-to-decision chain

6. Decision Gates

Gate 0: Use-Case Eligibility

QuestionPass conditionEvidence
Which decision will the forecast influence?Decision and owner documentedForecast Contract
Could the output affect funding, pricing, liquidity, ALM, FTP or board MI?Impact tier assignedUse Case Risk Tier
Is AI needed, or can deterministic rules / existing ALM process answer it?Alternatives recordedAlternatives note
What actions are prohibited?Explicit action boundaryAction Boundary Card
Who owns applicability and policy interpretation?Legal/Compliance/Model Risk owners namedGovernance routing record

Gate 1: Data Readiness

QuestionPass conditionEvidence
Are sources reconciled to GL / authoritative systems where needed?Reconciliation pass or issue acceptedReconciliation report
Is event time separate from available time?As-of design implementedData lineage record
Are product/segment/legal-entity mappings stable?Mapping versioned and approvedMapping catalog
Are restatements and corrections tracked?Change log existsRestatement log
Are sensitive data and customer-level signals governed?Data governance route completedData use record

Gate 2: Model And Assumption Design

QuestionPass conditionEvidence
Is there an honest baseline?Baseline model definedBaseline benchmark
Are outputs probabilistic where action risk requires it?Interval/quantile output availableForecast output spec
Is deposit beta/runoff segmented?Segment and scenario rationale documentedNMD Assumption Register
Are challenger models or expert overlays available?Challenger/overlay workflow activeChallenger report
Are limitations explicit?Model card and limitation log approvedModel risk pack

Gate 3: Stress And ALM Integration

QuestionPass conditionEvidence
Are scenarios versioned and approved?Scenario library entry approvedScenario Definition Card
Does liquidity ladder include contingent and operational flows?Full flow taxonomy implementedLadder specification
Do NII/EVE outputs use approved assumptions?ALM assumption linkage existsALM assumption trace
Is FTP impact calculated where incentives matter?FTP simulation integratedFTP impact report
Are legal-entity/currency constraints visible?Constraint layer enabledConstraint report

Gate 4: Release And Governance

QuestionPass conditionEvidence
Can ALCO challenge the output?Workbench shows drivers, alternatives, limitationsALCO pack sample
Are human overrides controlled?reason, approver, expiry and impact capturedOverlay record
Are actions dual-controlled?approval and execution separatedAction register
Can board MI be traced to evidence?metric contracts and lineage links existBoard MI lineage sample
Is monitoring live?performance, drift, data quality, issue aging activeMonitoring dashboard

7. Required Artifacts

ArtifactMinimum contents
Forecast Contractdecision use, grain, horizon, cadence, owner, outputs, allowed actions, prohibited use, evidence
Treasury Data Contractsource systems, fields, event time, available time, quality rules, reconciliation, retention
NMD Assumption Registerbeta, runoff, average life, repricing rate by product/segment/scenario, owner, approval
Scenario Definition Cardscenario type, horizon, rate path, runoff shock, funding shock, collateral haircut, operational assumption
Model Cardpurpose, design, data, performance, limitations, validation, monitoring, approved use
ALCO Workbench Specmetrics, scenarios, explanations, limits, action options, challenge fields
FTP Impact Reportproducts/business lines affected, funding cost, liquidity premium, contingent liquidity cost
Overlay Recordreason, scope, duration, approver, quantitative impact, expiry and post-review
Action Registertrigger, proposed action, approval forum, execution status, owner, closure evidence
Board MI Metric Contractdefinition, source lineage, threshold, owner, cadence, decision use, action link
Evidence Bindersource-to-decision graph, data snapshot, model/scenario version, committee record, action log

8. Data Contracts

8.1 Deposit Flow Contract

FieldDefinition
account_idstable account key or governed token
customer_group_idrelationship / household / commercial group, versioned
legal_entitybooking entity
product_typestandardized deposit taxonomy
event_timewhen balance-affecting event occurred
available_timewhen event was available to forecast system
amountsigned inflow/outflow amount
balance_afteraccount balance after event
rate_paidcurrent paid rate at event time
pricing_event_idlinked campaign / product rate change if applicable
channelbranch, online, wire, ACH, internal transfer, sweep
quality_statuspass, warning, excluded, reconciled

8.2 Cash-Flow Forecast Output Contract

FieldDefinition
forecast_idunique forecast run
forecast_contract_idgoverning contract
as_of_timedata cutoff
scenario_idbase or stress scenario
time_bucketintraday/day/week/month bucket
legal_entity / currencyliquidity transfer constraints
flow_typecontractual, behavioral, operational, contingent
p10 / p50 / p90probabilistic outputs
driver_refexplanation artifact
model_versionmodel and feature version
limitation_flagknown caveat for use

8.3 Decision And Action Contract

FieldDefinition
trigger_idlimit breach, warning, scenario result or committee request
proposed_actionfunding, pricing, collateral, portfolio, CFP, monitoring
action_boundaryread-only, recommend, committee approval, dual-control execution
owneraccountable decision owner
approval_forumTreasury, ALCO, Risk, Board, other approved forum
rationalestructured reason and evidence links
constraint_checklegal entity, currency, collateral, operational, policy
execution_statusproposed, approved, rejected, executed, closed
closure_evidenceexecution confirmation or reason for no action

9. Model And Scenario Selection

Use casePreferred model familyRequired controls
Daily cash positionrules + statistical / ML forecastdata freshness, cutoff, reconciliation, exception alerts
Intraday liquidityevent-driven forecast + payment rail calendarlow-latency monitoring, operational dependency, fallback
Retail NMD runoffsurvival / hazard, hierarchical, gradient boosting, probabilistic TSsegment logic, beta/runoff register, challenger, stress overlay
Commercial deposit concentrationconcentration graph + anomaly detection + relationship featurestop depositor lineage, group mapping, human review
Rate sensitivity / betaelasticity models, regime-aware regression, causal sensitivityrate campaign leakage check, competitor rate context
Loan prepaymentborrower behavior model + rate incentiveALM assumption approval, validation, scenario response
Commitment drawdownhistorical utilization + stress overlayscontingent liquidity evidence, expert challenge
ALM NII/EVEexisting ALM engine with AI assumption challengerassumption trace, NII/EVE bridge, model risk controls
FTP simulationrule/curve-based engine + scenario inputsmethodology transparency, business-line incentive review
Board MI narrativesource-grounded LLM summarizerapproved data only, citations, human review, no unsupported conclusion

Selection rule:

Use the simplest model that can support the decision, uncertainty, stress behavior,
explainability and evidence requirement.

10. Forecast-To-Action Matrix

SignalExample thresholdRequired actionOwnerEvidence
GreenForecast within normal band, data quality passnormal monitoringTreasurydashboard snapshot
Watchsegment runoff rising but within appetiteanalyst review and commentaryTreasury / ALMdriver bridge
AmberP90 runoff or survival horizon breaches internal watch thresholdALCO / Treasury Risk reviewALCO ownerscenario pack, challenge log
Redstress survival horizon below appetite or severe concentration movementmanagement escalation, CFP option reviewTreasurer / CROaction register, committee record
CFP triggerapproved contingency threshold reachedactivate approved CFP workflowauthorized management forumCFP action evidence
Data quality blocksource reconciliation fails for material inputblock board MI or label limitationData Owner / Technologyquality issue log
Model limitationmodel drift or interval coverage failurerestrict use, apply approved overlay, remediateModel Owner / Model Riskmodel issue record

Action principles:

  • A red signal without an action owner is not a control.
  • A model limitation without a use restriction is not transparent governance.
  • A human overlay without expiry becomes shadow methodology.
  • A board metric without source lineage is weak MI.

11. Evidence And Control Checklist

Control objectiveProduct controlEvidenceOwner
Forecast used for approved purposeForecast Contract enforced in UI/APIcontract id on every forecast outputPM / Treasury
Data is as-of correctevent_time and available_time capturedlineage and cutoff reportData Owner
Data quality is knowncompleteness/freshness/reconciliation rulesquality dashboard and issue logData / Technology
Deposit assumptions are governedNMD Assumption Registerapproval and version historyALM
Model is validated for intended usemodel card, validation and challengervalidation report and issue logModel Risk
Stress scenario is reproduciblescenario library with parametersscenario definition cardRisk / Treasury
AI explanation is groundedexplanation links to data/scenario/model refsdriver decomposition artifactProduct / Data Science
Committee challenge is realchallenge questions and responses capturedALCO / Risk minutesCommittee Secretary / Owner
Human overlay is controlledreason, approver, expiry, impactoverlay recordTreasury / Risk
Actions are authorizeddual control and limit checksaction registerTreasury Operations
Board MI is traceablemetric contracts and lineage linksboard MI lineage packMI Owner
Issues are closedaction owner and due dateremediation logRisk / Audit

12. Monitoring Metrics And KRIs

12.1 Forecast Performance

MetricWhy it matters
forecast error by horizon and segment看模型在哪里失效, 不只看全行平均
prediction interval coverage检查 P90 是否真的覆盖 tail behavior
directional accuracy around stress windows检查预警价值
calibration by product and legal entity防止某些实体/产品持续偏差
challenger disagreement rate衡量模型不确定性和有效挑战

12.2 Liquidity And ALM KRIs

MetricDecision use
stress survival horizon by scenarioliquidity appetite and CFP readiness
top depositor concentration movementconcentration risk and early warning
rate-sensitive deposit sharepricing and runoff vulnerability
beta / runoff assumption driftALM and FTP assumption review
contingent liquidity exposureCFP and FTP charge review
NII / EVE sensitivity bridgeALCO balance-sheet strategy
usable liquidity by legal entity/currencytransferability and constraint visibility
collateral encumbrance and haircut sensitivitystress funding capacity

12.3 Governance Metrics

MetricWhy it matters
model validation status by use caserelease and continued use
data quality exception agingMI reliability
human override rate and reason mixmodel trust, business change, shadow policy
action aging after amber/red signalmanagement responsiveness
evidence completeness scoreaudit and board readiness
expired assumption countALM discipline
unresolved model limitationsuse restriction and residual risk

13. Board MI Pack Design

Board pack should show decision-useful facts, not raw model outputs.

13.1 Minimum Board View

SectionContent
Executive liquidity statusgreen/amber/red, key change since last period, decision requested
Liquidity buffer and survivalbase and stress survival horizon, usable liquidity by constraint
Deposit behaviorbeta/runoff by key segment, top depositor concentration, digital/high-rate segment movement
ALM/IRRBBNII/EVE sensitivity and assumption changes
FTP / portfolio incentivesmajor products or business lines where incentives changed
CFP readinesstested capacity, action execution time, dry-run findings
AI/model riskmodel status, limitations, challenger disagreement, validation issues
Management actionsapproved, rejected, executed, overdue, residual risk owner

13.2 Board MI Tile

Topic: 30-day liquidity outlook under idiosyncratic deposit confidence stress

Status: Amber
Metric: stress survival horizon 43 days, down 11 days from prior quarter.
Drivers: P90 runoff increase in digital money-market and high-yield promotional cohorts; competitor-rate spread widened; no material change in top depositor concentration.
ALM impact: higher retention pricing improves runoff outcome but reduces projected NII; EVE sensitivity unchanged within appetite.
FTP impact: promotional deposit stability credit reduced; contingent liquidity charge increased for affected campaign balances.
Controls: challenger model confirms direction; primary model magnitude higher; 60-day expert overlay approved with expiry.
Management action: ALCO approved targeted retention for operational relationship customers and a CFP operational dry-run.
Evidence: data cutoff, model version, scenario id, overlay record, ALCO decision and action owner linked.
Decision requested: note amber status and challenge CFP dry-run completion timeline.

14. Tabletop Scenarios

Scenario A: Digital Deposit Run Triggered By Confidence Event

ElementDesign
Triggerunusual digital outflow, call center concern, media/news signal, top depositor inquiry
AI roleearly warning, segment contribution, P90 runoff forecast, scenario comparison
Human actionTreasury Risk review, ALCO escalation, CFP option assessment
Evidenceevent sources, customer segment movement, scenario version, action log
Failure to testAI overreacts to noisy NLP signal or misses concentration transfer

Scenario B: Rapid Rate Hike And Competitor Pricing Shock

ElementDesign
Triggermarket curve shift and competitor rate spread widening
AI rolebeta update, migration forecast, NII/liquidity trade-off
Human actionpricing committee and ALCO review
Evidencerate data source, pricing campaign linkage, beta assumption change
Failure to testbroad promotional pricing improves balances but destroys margin and attracts unstable funds

Scenario C: Payment Rail / Settlement Disruption

ElementDesign
TriggerACH/wire/card settlement delay or treasury platform degradation
AI roleintraday liquidity impact, operational cash-flow reroute options
Human actionincident command, treasury operations, BCP invocation
Evidencepayment event logs, BCP record, liquidity source availability
Failure to testliquidity exists on paper but cannot be moved operationally

Scenario D: Collateral Haircut And Wholesale Funding Stress

ElementDesign
Triggermarket liquidity shock, securities haircut change, rollover pressure
AI rolehaircut sensitivity, funding source availability, survival horizon
Human actioncollateral strategy, funding diversification review
Evidencecollateral inventory, encumbrance, haircut scenario, funding action feasibility
Failure to teststress assumes asset sale that is not operationally or market-feasibly executable

Scenario E: Model Drift During Benign Period

ElementDesign
Triggerlow error in aggregate but poor interval coverage for key segment
AI rolemonitoring flags calibration gap and challenger disagreement
Human actionrestrict use for affected segment, approve overlay, start remediation
Evidencebacktest, interval coverage, model limitation log, overlay expiry
Failure to testgood portfolio-level accuracy masks material risk pocket

15. Implementation Roadmap

Phase 1: Foundation, 0-30 Days

WorkstreamDeliverable
Scopeprioritized use cases: daily liquidity, NMD runoff, ALCO stress pack, board MI
Governanceforecast contracts, action boundaries, RACI, committee routing
Datasource inventory, data contracts, lineage gaps, reconciliation requirements
Baselinesimple rules/seasonal baseline for key cash-flow and deposit forecasts
Evidencedraft evidence graph and board MI metric contracts

Phase 2: Pilot, 31-60 Days

WorkstreamDeliverable
Deposit behaviorsegment beta/runoff prototype with uncertainty
Liquidity laddercontractual + behavioral + operational + contingent flow view
Scenariobase, idiosyncratic, market-wide and combined scenario cards
ALCO workbenchscenario comparison, drivers, limits, override capture
Model riskmodel cards, validation plan, challenger baseline, monitoring metrics

Phase 3: Controlled Release, 61-90 Days

WorkstreamDeliverable
Production dataautomated lineage, quality gates, reconciliation and data issue workflow
Decision workflowaction register, committee challenge log, dual-control integration
ALM/FTPNII/EVE bridge and FTP simulation for selected products
Board MItraceable board tile and evidence export
Assurancetabletop scenario, dry-run, issue remediation and release sign-off

Phase 4: Scale

WorkstreamDeliverable
Portfoliomulti-entity, currency, product and business-line expansion
Stressreverse stress, intraday stress, operational stress and CFP dry-run library
GovernanceAIMS integration, internal audit testing, continuous monitoring
Optimizationscenario-aware portfolio/pricing/funding option analysis under constraints

16. Anti-Patterns

Anti-patternWhy it failsBetter practice
"AI predicts deposits" as project scope预测没有决策边界use-case-specific forecast contracts
Single dashboard for Treasury, ALM, Board不同决策需要不同 evidence grainrole-specific MI from same data product
Model-first delivery忽略 committee, actions, controlsforecast-to-action workflow first
P50 forecast as liquidity truthtail risk is the pointquantiles, stress and survival horizon
Deposit beta as one scalarhides segment, rate and relationship behaviorsegmented assumption register
Manual ALCO packno reproducible evidencemetric contracts and lineage
LLM-generated narrative without source groundingplausible but unauditablegrounded summary with evidence links and human review
No challenger modelno effective challengehonest baseline and challenger dashboard
Unlimited human overridesshadow methodologycontrolled overlay with reason, expiry and post-review
CFP actions not testedplan may not execute under stressdry-run evidence and operational readiness

17. Interview And Portfolio Language

Q1: What would you build first for AI treasury liquidity forecasting?

30 秒版本:

I would build the forecast contract and evidence spine first, not the most advanced model. The first release should define decision use, grain, horizon, data lineage, baseline forecast, deposit behavior assumptions, scenario definitions, ALCO challenge workflow and board MI metric contracts.

Q2: How do you prevent AI from becoming a black box in ALM?

30 秒版本:

I separate AI assumptions from ALM engines. AI can challenge NMD beta, runoff, prepayment and scenario assumptions, but approved ALM assumptions remain versioned governance objects. ALCO sees driver bridge, challenger comparison, uncertainty, limitations and override records before using them in NII/EVE decisions.

Q3: What is a good evidence pack for stress testing?

30 秒版本:

A good pack includes scenario definition, data cutoff, model and feature versions, assumptions, overlays, liquidity ladder output, ALM impact, FTP impact, limit status, committee challenge, management actions and post-event backtesting. The key is source-to-action traceability.

Q4: How would you connect FTP to AI forecasts?

30 秒版本:

I would use AI to estimate behavior and stress sensitivity, then feed approved assumptions into FTP. Stable operational deposits may receive funding benefit only when stability evidence supports it; promotional or rate-sensitive balances receive lower stability credit and potentially higher contingent liquidity charges. FTP becomes the incentive layer for liquidity risk.

Q5: How do you phrase this for senior stakeholders?

30 秒版本:

The goal is not to automate Treasury. The goal is to give Treasury, ALCO and the board a faster and better-evidenced view of liquidity risk, assumption changes and management options, while preserving human decision rights and auditability.


18. Final Checklist

QuestionGood answer
Can every forecast output identify its contract?yes, forecast_contract_id is attached
Can the team replay a board liquidity metric?yes, metric -> data -> model -> scenario -> decision -> action
Are deposit assumptions segmented and approved?yes, NMD register has owner, version and evidence
Are stress scenarios reproducible?yes, scenario cards are versioned
Can ALCO challenge model outputs?yes, drivers, alternatives and limitations are visible
Are overrides controlled?yes, reason, impact, approver, expiry and review exist
Are actions authorized and tracked?yes, action register links to trigger and evidence
Is model risk live after release?yes, monitoring, drift, backtest, issue and limitation tracking run continuously
Is exact regulatory applicability owned by the right functions?yes, Legal/Compliance/Model Risk own formal applicability and conclusions

Memory card:

AI Treasury Playbook =
  forecast contracts
  + as-of data lineage
  + deposit behavior assumptions
  + liquidity ladder
  + ALM / IRRBB / FTP integration
  + stress evidence
  + human committee
  + board MI
  + continuous model risk monitoring.