AI Governance Treasury ALM Banking: Interest Rate Risk AI, Liquidity Forecasting AI, IRRBB Governance, and FTP Model Risk

IRRBB behavioral models, LCR/NSFR forecasting AI, FTP curve construction, and investment portfolio AI — where treasury AI governance failures show up directly in regulatory capital, not just in customer complaints.

Treasury AI governance gets less attention than credit or fraud AI, which is backwards, because treasury models have no buffer between a bad assumption and a bad number. A credit model that's 5% less accurate costs money gradually, across thousands of individual decisions. An IRRBB model that understates NII sensitivity at a 200bps shock produces an understated Pillar 2 capital add-on directly — one number, wrong, feeding straight into the capital stack. When behavioral assumptions for non-maturity deposits drift from reality, or an FTP model misprices the liquidity premium on five-year mortgages, the error isn't a tail risk sitting in a risk register. It's a running miscalculation compounding every day until someone catches it.

What makes governance here genuinely hard — not just paperwork-hard — is that treasury AI sits at the intersection of three different demands simultaneously: the regulatory framework (Basel IRRBB standards, RBI's IRRBB and LCR guidelines), the mathematics (stochastic rate models, option-adjusted spread, behavioral finance for deposits), and the operational reality of getting a number in front of ALCO that a non-quant committee member can actually challenge. A governance program that nails the regulatory checklist but produces a model the treasury head can't explain to ALCO hasn't actually solved the problem.

IRRBB AI: Behavioral Assumptions Have to Survive a Supervisor's Challenge, Not Just a Backtest

Basel's BCBS d368 standards require NII and EVE sensitivity measurement across six prescribed rate scenarios plus internal stress scenarios, and AI used for that measurement inherits capital-model-level governance requirements. The behavioral assumptions are where this gets tested first: NMD stability and prepayment assumptions need documentation and historical validation that holds up under supervisory challenge, and RBI's IRRBB guidance sets floors and caps on these parameters that an AI-generated assumption has to land inside, not just approximate.

Scenario consistency is the check most teams skip until a validator catches it. If an AI model shows NII and EVE moving in the same direction under an upward rate shock — unusual for a typical retail balance sheet — that's not a quirky output, it's a validation finding waiting to be written up. And none of this matters if the treasury head can't stand in front of ALCO and explain, in plain terms, why the model says what it says: a board setting IRRBB risk appetite against a number nobody in the room can unpack isn't actually exercising risk appetite.

Liquidity AI: LCR, NSFR, and the Trap of Optimizing the Ratio Instead of Measuring It

LCR AI has one failure mode that looks like success until an examiner digs in: a model that finds creative interpretations of HQLA eligibility or haircuts to improve the forecast ratio is optimizing the wrong thing. HQLA classification, haircuts, and outflow rates have to match Basel III and RBI's LCR definitions exactly — there's no AI-discovered edge case that beats a precise definition.

Stress scenario calibration is the second trap: RBI's LCR stress scenarios combine idiosyncratic and market-wide stress, and a behavioral model calibrated to optimistic central estimates instead of the conservative assumptions the framework requires will pass a backtest and fail a stress review. LCR and NSFR also need to come out of a coupled model, not two separate ones that can quietly diverge and create an inconsistency a supervisor will ask about. And intraday liquidity AI has a different accuracy bar entirely — RTGS and NEFT clearing run on sub-hour decision cycles, so an intraday model validated only on end-of-day accuracy is validated for the wrong problem.

FTP AI: The Curve Has to Be Auditable, Not Just Optimized

Funds transfer pricing touches every business line's P&L, which means an opaque FTP curve doesn't just create a model risk issue — it creates a cross-business-line trust issue. The curve has to be built from observable market rates (G-Sec yields, swap rates, credit spreads) through a documented methodology; an AI optimization that pulls in unobservable or circular inputs produces a curve nobody can audit, which defeats the entire purpose of FTP as an internal pricing mechanism.

Liquidity premium AI needs to track the bank's actual current funding cost and position, not a smoothed historical average — in stress conditions, a lagging liquidity premium hides real funding cost from the business lines pricing loans against it, right when they need the real number most. For India specifically, AI assisting MCLR computation has to track actual marginal funding cost, not a competitively optimized number — RBI's MCLR framework is explicitly cost-based, and an AI model tuned for competitive positioning instead of cost accuracy is non-compliant by design, not by accident. And an FTP rate a business line head can't get explained to them is a rate they can't manage against — which quietly breaks ALCO's ability to manage the bank's aggregate interest rate position.

Investment Portfolio AI: HTM/AFS Classification and G-Sec Trading

AI assisting HTM versus AFS classification has to apply RBI's criteria precisely — HTM is permitted for SLR securities up to a prescribed ceiling (currently 23% of NDTL), and a misclassification here has both regulatory capital and P&L consequences, not just a model accuracy footnote. G-Sec trading AI used for duration management needs to comply with RBI's government securities rules and SEBI's G-Sec market guidelines, with governance covering pre-trade limit checks, OMS integration, and a genuine audit trail for every AI-assisted decision. Derivative overlay AI used in IRRBB hedging needs hedge-accounting documentation that satisfies IndAS 109's hedge effectiveness testing — a hedge that works operationally but can't produce the accounting documentation isn't a complete governance solution.

India Context: Where RBI's Rules and India's Balance Sheets Create Specific Obligations

RBI's IRRBB guidance layers India-specific adjustments onto the Basel baseline — floors on NMD core-deposit stability assumptions, prescribed shock scenarios reflecting India's own rate-transmission dynamics, and a requirement that IRRBB risk appetite be board-approved and reviewed at least quarterly, not annually. A bank using AI for behavioral modeling needs its outputs to land inside those prescribed bounds and needs to be ready to explain the modeling approach under RBI's SREP review — "the model said so" has never satisfied a SREP examiner and won't start now because the model is AI.

Public sector bank treasury has a governance wrinkle private banks don't carry: government-directed lending — priority sector targets, MSME mandates — shifts the balance sheet's repricing profile in ways an AI model trained on historical PSB data won't anticipate on its own. A sudden directed-lending push into agriculture changes the banking book's interest rate sensitivity in ways that need to be modeled explicitly, not assumed away. And India's payment infrastructure creates an intraday liquidity challenge with no real international parallel: UPI's transaction volume produces distinct intraday peaks, and RTGS queuing dynamics are India-specific — an intraday liquidity model trained on international behavioral datasets simply hasn't seen this pattern before.

Action Checklist: Treasury AI Governance That Survives a SREP Review

1. Validate behavioral assumptions against RBI's prescribed floors and caps — don't just backtest against your own historical data.
2. Check NII/EVE scenario consistency explicitly across all six Basel shocks; investigate any same-direction movement before a validator does.
3. Decouple-test LCR and NSFR outputs to confirm they come from (or are reconciled against) the same underlying model — don't let them diverge silently.
4. Require that every FTP curve input be traceable to an observable market rate — no AI-optimized input that can't be explained to a business line head.
5. Model directed-lending and UPI/RTGS intraday effects explicitly rather than assuming historical training data already captures India-specific balance sheet dynamics.