What AI models are used in IRRBB measurement and how does MRMF 2026 classify them?

RBI's IRRBB Guidelines (February 17, 2023, DOR.MRG.REC.14/21.04.098/2023-24) require banks to model behavioral aspects of their banking book — deposits without contractual maturity (current and savings accounts), loan prepayments, and optionality in fixed-rate products. These behavioral models require estimates of customer behavior under different interest rate environments, and AI/ML models are increasingly used to derive these estimates from historical data. Four AI model types are used in IRRBB measurement: (1) Non-maturity deposit (NMD) behavioural models: AI models that forecast the portion of current/savings account balances that will remain with the bank under various rate scenarios — critical for EVE and NII sensitivity calculations; (2) Loan prepayment AI: models forecasting prepayment rates for fixed-rate home and auto loans under rising and falling rate environments — a significant source of IRRBB in India's large retail banking book; (3) Yield curve forecasting AI: ML models that generate interest rate scenario distributions beyond the RBI-prescribed shock scenarios, used in internal capital adequacy assessments; (4) EVE/NII simulation AI: integrated models that calculate Economic Value of Equity and Net Interest Income sensitivity by combining behavioral assumptions with rate shock scenarios. Under RBI MRMF 2026, all four are classified as quantitative models used in material risk calculations and require model inventory registration, independent validation by an internal model validation unit or external validator, and board-level reporting on model risk exposure.

What is the dual IRRBB-MRMF regulatory overlap and why does it create unique compliance overhead for India banks?

India's IRRBB AI models are dual-regulated in a way that has no parallel in EU or US banking frameworks. The RBI IRRBB Guidelines (February 2023) require that banks validate their behavioral models for IRRBB measurement purposes — the IRRBB validation focuses on the accuracy of the behavioral assumptions (do the NMD model's predicted runoff rates match historical observed runoffs?) and the appropriateness of the rate shock scenarios applied. The RBI MRMF (June 2026) requires that quantitative models used in material risk calculations undergo independent model validation — the MRMF validation focuses on model development soundness, data quality, conceptual soundness, and ongoing performance monitoring against statistical benchmarks. These are different validation objectives, requiring different validation methodologies, producing different validation reports for different regulatory audiences. A bank that conducts a single validation and presents it to both the IRRBB examination team and the MRMF examination team may find that the IRRBB-focused validation does not satisfy MRMF's statistical performance benchmarks, or that the MRMF-focused validation does not address the IRRBB-specific behavioral assumption verification. The compliance overhead is separate validation records for each regulatory dimension — IRRBB validation records for RBI's market risk examination and MRMF validation records for RBI's model risk governance examination. India's banks face this dual-track validation obligation that neither EU CRR/CRD banking supervisors (who have their own IRRBB framework) nor US Federal Reserve IRRBB supervisors impose simultaneously with a comprehensive model risk management standard.

When does AI scenario generation outside RBI IRRBB shock bounds require board approval?

The RBI IRRBB Guidelines prescribe six interest rate shock scenarios for EVE and NII sensitivity measurement: parallel shock up, parallel shock down, short rate shock up, short rate shock down, steepener shock, and flattener shock. Banks must compute IRRBB exposures for all six scenarios and report them to RBI. For internal risk management and internal capital adequacy assessments (ICAAP under Pillar 2), banks may use additional scenarios beyond the six prescribed shocks — including AI/ML-generated scenarios derived from historical yield curve dynamics, macroeconomic forecasting models, or neural network-based term structure models. An AI yield curve forecasting model that generates scenarios outside the RBI-prescribed shock bounds — for example, scenarios involving simultaneous short-rate and long-rate movements in directions not covered by the six prescribed shocks — is generating scenarios for use in material capital adequacy calculations. Under MRMF's 'material model change' provisions, deploying a new scenario generation methodology in a model that feeds ICAAP requires documented approval. The specific IRRBB governance obligation: if an AI scenario generation model produces scenarios that, when used in ICAAP, result in materially different capital adequacy outcomes than the six prescribed RBI scenarios, this represents a material model decision requiring board-level approval and documentation in the bank's ICAAP submission to RBI. RBI's SREP review of ICAAP includes assessing the appropriateness of scenario assumptions — an AI-generated scenario that a bank cannot defend under supervisory scrutiny creates SREP exposure.

How does the November 2025 RRB ALM extension create new third-party AI governance obligations?

RBI extended IRRBB measurement obligations to Regional Rural Banks (RRBs) through its November 2025 ALM circular. India has 43 Regional Rural Banks (post-amalgamation). Most RRBs do not have internal model development capacity — they rely on their sponsor banks or third-party vendors for ALM and IRRBB models. The November 2025 extension means that RRB boards are now accountable for IRRBB models they almost universally did not develop and may not have had previously validated. Three AI governance implications: (1) Sponsor bank model governance: RRBs using their sponsor bank's IRRBB models must ensure those models have been independently validated on the RRB's specific balance sheet composition — not just the sponsor bank's consolidated portfolio. RRB deposit mix, NMD runoff characteristics, and loan prepayment behavior differ materially from large commercial banks; a model validated on the sponsor's balance sheet may not be appropriate for the RRB; (2) Third-party vendor liability: RRBs using ALM software vendor models face MRMF Section 7 obligations — the vendor's model must be in the RRB's MRMF inventory (a new MRMF inventory for most RRBs), independently validated, and board-reported. Many ALM software vendors have not yet produced MRMF-compliant model documentation for their India banking product customers; (3) Capacity requirement: RRBs with no internal model validation unit must either build one or commission external validation — a market opportunity for model risk governance services but a compliance obligation that begins to run immediately under the November 2025 circular.

AI Governance · RBI IRRBB · ALM · Treasury · MRMF · India BFSI

RBI IRRBB AI Model Governance: Interest Rate Risk in the Banking Book, ALM Behavioural Models, and MRMF Obligations for India Banks 2026

RBI's IRRBB Guidelines (February 2023, DOR.MRG.REC.14/21.04.098/2023-24) require behavioural models for deposit repricing, prepayment, and optionality. These AI models are classified as quantitative models under MRMF 2026 and require independent validation and board reporting. IRRBB AI is dual-regulated — IRRBB validation and MRMF validation are separate regulatory obligations producing separate evidence packages. AI scenario generation outside RBI shock bounds requires board approval. Capital adequacy understatement from AI error is a Pillar 2 breach. Zero global competitor coverage.

Zytra Research Team · October 2026 · 13 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

Interest Rate Risk in the Banking Book (IRRBB) is among the most technically demanding risk management domains in banking. India's IRRBB regulatory framework, consolidated in the RBI Guidelines on Interest Rate Risk in the Banking Book (DOR.MRG.REC.14/21.04.098/2023-24, dated February 17, 2023), requires banks to measure and manage the risk that interest rate movements will affect the economic value and earnings of their banking book — the retail deposits, fixed-rate loans, and ALM instruments that constitute the bulk of an Indian commercial bank's balance sheet.

Measuring IRRBB accurately requires behavioral models: estimates of how depositors will behave when rates change (will they move current account balances to fixed deposits?), how borrowers will prepay when rates fall (will home loan borrowers refinance?), and how the bank's own optionality (callable bonds, early repayment penalties) will be exercised. AI and ML models are the current state of the art for these behavioral estimates — trained on historical transaction data, they can capture non-linear rate-behavior relationships that traditional econometric models miss.

MRMF 2026 classifies these AI behavioral models as quantitative models used in material risk calculations. They must be governed under MRMF. But they are simultaneously subject to RBI's IRRBB framework requirements. This dual regulatory overlap creates a compliance overhead that is unique to India — no global banking jurisdiction imposes simultaneous comprehensive IRRBB and model risk management governance obligations on the same AI models.

AI Models in IRRBB Measurement: What Is In Scope

Four AI model categories are in scope for IRRBB AI governance in India:

Non-Maturity Deposit behavioural models: Current Account and Savings Account deposits have no contractual maturity. Banks must model how these deposits behave under rate changes — what fraction will remain "core" (stable, long-duration), what fraction will reprice quickly as depositors move funds. NMD models drive EVE sensitivity calculations; an AI model that overestimates the core stable fraction understates IRRBB EVE sensitivity.

Loan prepayment models: India's large fixed-rate retail lending book (home loans, auto loans) creates IRRBB through prepayment optionality — when rates fall, borrowers prepay and refinance, compressing the bank's interest income. AI prepayment models estimate the rate-sensitivity of prepayment in each rate environment. Underestimating prepayment probability understates IRRBB NII sensitivity.

Yield curve scenario generation AI: Banks use AI to generate internal rate scenarios for ICAAP beyond the six RBI-prescribed shocks — incorporating macroeconomic forecasts, regime-switching models, and neural network term structure models. These AI-generated scenarios feed capital adequacy calculations.

EVE/NII integrated simulation AI: Integrated platforms that compute EVE and NII sensitivity by combining behavioral model outputs with rate shock scenarios and balance sheet projections. The integrated model's accuracy depends on the quality of each behavioral model component.

Dual IRRBB-MRMF Validation: Two Records, Two Reviewers

RBI IRRBB examinations and MRMF examinations are conducted by different teams with different objectives. An IRRBB examination focuses on whether the bank's IRRBB measurement is conceptually sound, the behavioral assumptions are justified by historical data, and the six prescribed scenarios are correctly computed. An MRMF examination focuses on whether the quantitative models have been independently validated, are being monitored for drift, and are properly inventoried and board-reported.

A single validation report cannot satisfy both examination teams. The IRRBB-focused validation report may confirm that the NMD model's core deposit assumptions are historically justified but not address the MRMF statistical performance benchmarks (PSI, back-testing accuracy, out-of-sample validation). The MRMF-focused validation may confirm the model's development soundness but not address the IRRBB-specific question of whether the behavioral assumptions are appropriate for the bank's specific deposit mix.

India banks must maintain separate validation records: IRRBB validation documentation for the market risk examination, and MRMF validation documentation for the model risk governance examination. Both must be current, because both examination teams may request them on any examination cycle.

AI Scenario Generation: The Board Approval Trigger

RBI's IRRBB Guidelines prescribe six shock scenarios. For ICAAP (the bank's internal capital adequacy assessment), banks may use additional scenarios. When AI models generate those additional scenarios, and the AI-generated scenarios produce materially different capital adequacy outcomes than the six prescribed scenarios, MRMF's material model decision framework applies.

The governance question: which AI-generated scenario outcomes require board-level review? A practical answer: when AI scenario generation produces capital requirement estimates that deviate by more than a defined threshold from the six prescribed scenarios, that deviation should be reviewed by the board's risk committee before inclusion in ICAAP. The board approves the ICAAP; if AI models are driving material capital adequacy decisions through non-prescribed scenarios, the board's approval of ICAAP should be informed by those AI model outputs.

RBI's SREP review will assess ICAAP assumptions, including the appropriateness of stress scenarios. An AI-generated scenario that the bank cannot defend methodologically to RBI supervisors creates SREP exposure. MRMF's documentation requirement for AI scenario generation models must include the scenario generation logic, validation on historical yield curve behavior, and sensitivity analysis showing the capital impact of scenario parameter choices.

Pillar 2 Consequences of AI IRRBB Model Failure

AI model failures in IRRBB measurement have consequences that other AI governance domains do not: they propagate into regulatory capital calculations. A NMD behavioral model that systematically overstates core deposit stability understates IRRBB EVE sensitivity. EVE sensitivity feeds the bank's Internal Capital Adequacy Assessment Process (ICAAP). An understated ICAAP capital requirement represents an understatement of the bank's Pillar 2 capital buffer.

Under RBI's SREP framework, a Pillar 2 capital buffer understatement can result in RBI requiring the bank to hold additional capital above its Pillar 1 minimum — a direct financial consequence that is larger than any MRMF examination finding in isolation. The AI model failure translates into capital adequacy risk.

This creates a specific AI governance priority: IRRBB behavioral models must be treated as systemically important models within the MRMF framework, receiving the most rigorous validation, monitoring, and board oversight attention — not because they fail customers directly, but because their errors feed regulatory capital calculations with consequences on the bank's capital adequacy position.

November 2025 RRB Extension: Third-Party AI Governance at Scale

RBI's November 2025 ALM circular extended IRRBB measurement obligations to Regional Rural Banks. India's 43 RRBs are now required to measure and govern IRRBB. Most RRBs lack internal model development capacity — they will use their sponsor bank's models or third-party ALM software vendor models.

For MRMF compliance, each RRB must have its IRRBB models in its MRMF model inventory, independently validated on the RRB's own balance sheet composition, and board-reported. Sponsor bank models or vendor models applied without RRB-specific validation do not satisfy these requirements. The November 2025 extension creates a significant demand for IRRBB model governance services across India's RRB ecosystem — and a compliance clock that began running immediately on publication of the circular.

IRRBB AI Governance Checklist

☐ IRRBB AI model inventory: NMD, prepayment, yield curve, and EVE/NII simulation AI models all in MRMF model inventory with model cards; material model designation for ICAAP-feeding models
☐ Dual validation records: IRRBB-focused validation (behavioral assumption justification, scenario correctness) and MRMF-focused validation (statistical performance, PSI, out-of-sample) maintained separately
☐ Independent validation unit: IRRBB AI models validated by team independent of treasury/ALM; for smaller banks without internal MVU, external validation commissioned
☐ AI scenario generation board approval: process for identifying when AI-generated scenarios materially deviate from prescribed shocks; board risk committee review before ICAAP inclusion
☐ ICAAP AI documentation: AI model outputs used in ICAAP documented with model version, validation status, and sensitivity analysis; board-approved ICAAP records include AI model documentation
☐ Pillar 2 model risk capital: AI model uncertainty in IRRBB measurement reflected in ICAAP capital buffer; model risk contribution to IRRBB capital requirement documented
☐ PSI monitoring for behavioral models: ongoing monitoring of NMD and prepayment model performance against actual behavior; model update trigger when behavioral drift detected
☐ RRBs: IRRBB AI in RRB MRMF inventory; validated on RRB-specific balance sheet (not sponsor bank composite); third-party vendor model documentation access confirmed
☐ SREP readiness: IRRBB AI model documentation package (MRMF card, IRRBB validation, scenario generation rationale) ready for SREP examination on request
☐ Annual revalidation: IRRBB behavioral AI models revalidated annually minimum; event-triggered revalidation on material rate environment changes (RBI repo rate change above threshold)