Why most bank balance sheet optimization runs on judgment, not math

Portfolio optimization is a more familiar tool in asset management than in commercial banking. Not because every asset manager runs a mean-variance solver as the decision engine, many do not, but because the technique is at least in the working vocabulary and, where it is used, it is used to understand a single risk-return tradeoff across one dimension: market risk on total return. The objective is intelligible, the constraints are short, the result lands in language the investment committee can argue about.

Commercial banking faces the same conceptual problem with more dimensions, and the literature on how to solve it is mature. Lubinska 2020 publishes a fully specified funding-mix optimization with worked case studies on a multi-billion-dollar UK bank. Sodhi 2005 set out the multi-period stochastic taxonomy in

Operations Research two decades ago. Adam 2008 supplies the FTP architecture that makes the optimization tractable in production. Dermine 2003, Van Deventer 2013, and Choudhry 2018 supply the value-creation hurdle, the safety-zone framework, and the governance scaffolding. The Financial Risk Academy Building an ALM and Balance Sheet Optimization Model course ships a complete end-to-end implementation any practitioner can run on a laptop. The mathematics is settled. Why does almost no bank run it as the engine?

The answer is not about mathematics. It sits in four places: the behavioral, contract-driven nature of the bank balance sheet itself; the multi-dimensional constraint stack whose binding corner rotates with the regime; the four-silo treasury architecture that splits expertise across IRRBB, Liquidity, Capital, and Investment Portfolio specializations; and the gap between what a multi-dimensional solver outputs and what ALCO can connect, trust, and defend. A two-dimensional objective with a flat constraint corner is intelligible to the people accountable for the decision. The same machinery with eight dimensions, regime-dependent corners, and behavioral inputs the bank cannot control is not, particularly when each dimension is owned by a different team running a different model on a different platform with no shared layer that lets ALCO read across them.

The realistic form of balance sheet optimization in 2026 is a scenario-bounded playbook of pre-approved actions, and agentic AI is what finally makes that playbook usable in practice. The shift is not that AI solves the optimization problem. The shift is that AI synthesizes the expertise sitting in the four silos, encodes ALCO's rules and goals in a form the agent can iterate against, and runs the strategic menu across the constraint stack at a cadence the institution can actually use. AI does not replace the bank's experts. It connects them.

The frictionless illusion

The asset manager's world is close to frictionless. Positions sit on public exchanges, settle in seconds, reallocate at the click of an execution venue. If the Markowitz solver says sell muni bonds and buy tech equities, the trade completes intraday. The model output is the management action.

The bank's balance sheet does not work that way. A bank is not a portfolio of liquid claims; it is a collection of illiquid, non-linear retail and commercial contracts whose evolution depends on the behavior of millions of clients (Lubinska 2020). The deposit base depends on branch infrastructure, marketing, pricing, and client inertia. The loan book is bounded by macro credit demand, underwriting standards, and competitive position. The legacy back-book amortizes against contractual schedules and behavioral prepayment; the new front-book depends on what walks in the door. Adam 2008 makes the same point from the FTP side: every deal commits the bank to a multi-year IRR and liquidity exposure no solver can unwind by clicking sell, which is why each contract's transfer price is fixed at origination for life (Adam 2008).

An optimizer that prescribes a 60 percent mortgage share when the current share is 20 percent is not giving the bank an instruction. It is describing a counterfactual the bank cannot reach. Loan originations cannot be conjured by an LP; deposit volumes cannot be raised by a Lagrange multiplier. The frictionless model meets the unfrictionless balance sheet, and the recommendation arrives at ALCO with no path to execution.

The tri-lens stack

The second structural difference is dimensional. The asset manager runs one risk lens: market risk on total return. The bank runs three concurrent lenses, each with its own metric set, regulator, and constraint corridor. They do not jointly optimize, and the binding corner rotates with the regime.

IRRBB sits on the earnings and value side. The NII view measures how interest revenue on loans and interest expense on deposits reprice over a 12-month to 3-year horizon (OCC 2020). The EVE view treats the entire book as a portfolio of cash flows discounted to present value, measured against the supervisory outlier test corridor capped at 15 percent of Tier 1 across the six standardized shocks (BCBS 2016; EBA 2022). The two views conflict at the constraint corner. A bank can insulate short-term NII against falling rates by locking in long-dated fixed assets, but doing so increases asset duration and exposes EVE to a sharp loss if rates spike. NII immunization and EVE immunization are different mathematical conditions and cannot in general be solved simultaneously (Dermine 2003; Van Deventer 2013).

Liquidity sits on the survival side. The LCR floor over a 30-day combined stress (BCBS 2013) and the NSFR floor over a 12-month horizon (BCBS 2014) require the bank to hold HQLA and stable funding. HQLA yields meaningfully less than commercial lending. The bank that minimizes HQLA in pursuit of NIM breaches the LCR; the bank that maximizes HQLA in pursuit of liquidity comfort drags NIM into the floor.

Capital sits on the solvency side. Basel III binds the bank through CET1, Tier 1, and total capital ratios, the leverage ratio, and the concentration caps under BCBS 2013 paragraph 196 (the funding-concentration monitoring tool flags any connected counterparty above 1 percent of total liabilities). Capital constraints trigger discrete regulatory action at named thresholds.

Three lenses, three sets of metrics. The same dollar of balance-sheet capacity cannot satisfy all three at once. The binding lens rotates: in a rising-rate environment with a stable franchise, EVE binds; in a stressed-funding environment with uninsured-deposit flight, LCR binds; in a credit-cycle downturn, capital binds. A solver run on a single point estimate of which lens is at the corner produces a recommendation that is precise for that regime and infeasible for the next. What the literature did not settle is which corner the institution should sit at. That is an ALCO position, not a solver output.

The illusion of control

Even with the constraint stack stable, the optimizer cannot move the balance sheet to the position it prescribes. The asset manager controls the portfolio composition; the bank does not. Loan growth is bounded by macro credit demand and underwriting capacity; deposit volumes depend on client decisions made for reasons that have nothing to do with the solver's dual variables (Drechsler, Savov, Schnabl 2021; Drechsler et al. 2026). The standard LP or QP formulation, run without behavioral envelopes, returns corner solutions: zero out a loan category, double the deposit base, shift fifteen billion in HQLA composition in a single stage. None of these is executable.

The practical fix is the global turnover constraint, ∑i∣xt,i−xt−1,i∣≤T, where T is the institution's operational throughput per period (Financial Risk Academy). The constraint is a regularizer that keeps the solver from prescribing flows the bank cannot execute. It is also a confession that the optimization does not naturally produce implementable output. Coulier et al. (2025), working on confidential ECB supervisory data covering 67 euro-area Significant Institutions, document the same point from the calibration side. During the 2022-2023 tightening cycle, sensitive-NMD banks did not update their behavioral maturity assumptions; the average NMD maturity counterintuitively increased by 55 days because banks lost short-bucket deposits while retaining long-bucket allocations, without any model update. An optimization at those banks would have been running on stale coefficients in the direction opposite the realized regime.

The same point holds across the input stack. The IRRBB constraint feeds on six behavioral models: deposit beta, deposit decay, NMD repricing schedule, prepayment speed, draw rates, and option exercise. A point estimate of any one, fed into the LP, produces a precise recommendation; the same recommendation run at the 25th and 75th percentile of the empirical distribution can span billions of dollars of notional swap. I worked through the variance for the deposit-stability case in Stable, Structural, Core: One Word, Several Conditions (May 2026): the same NMD parameter, read against three different supervisory conditions, yields three different correct numbers, and the choice between them is an ALCO position rather than a modeling refinement.

The silo and the ALCO education gap

The third structural difference is organizational. The asset manager's portfolio sits on one platform, with one data model, run by one team against one objective. The bank's balance sheet sits across four operational silos.

The IRRBB desk runs a cash-flow modeling platform calibrated to the deposit-decay model and the NII corridor. The Liquidity desk runs a separate platform calibrated to LCR runoff factors and the contingency funding plan. The Capital team runs an RWA tracker and the regulatory reporting stack. The Investment Portfolio team runs a securities-portfolio optimizer over total return. Four desks, four platforms, four sets of behavioral assumptions, four data lineages. Cross-silo coordination happens by people, not by systems. A defensive IRRBB swap program can drain HQLA through margin calls before the Liquidity desk knows the trade went on. A Liquidity-driven shift toward shorter HQLA tenors can swing the IRRBB EVE corridor without the IRRBB desk having seen the trade.

SR 26-2, the April 2026 update to the FRB-OCC-FDIC model risk management guidance, recognizes aggregate model risk as the dependency structure across models that share assumptions, data, and methodologies (FRB, OCC, FDIC 2026). An ALM stack that feeds deposit beta, decay, NMD repricing, prepayment, draw rates, and option exercise into separate optimization platforms is the canonical aggregate-risk case. The deposit-side version is the cleanest example: the same word, read by three desks against three different conditions, produces three different correct numbers, and the gap is invisible until the IRR hedge and the contingent liquidity hedge end up on the same dollars.

ALCO sits at the apex and is supposed to negotiate the trade-offs the silos cannot see. The committee is composed of brilliant operators (the CEO, the CFO, the CRO, the Treasurer, the heads of business lines), each carrying deep commercial-banking experience. The expertise is asymmetric, by design (Choudhry 2018). The head of retail lending does not need to read a stochastic yield-curve simulation to authorize a billion-dollar funding shift; she needs a defensible rationale that maps the trade to the strategy she signed up for. A solver output of the form "the dual on the NSFR constraint is 32 basis points, the policy recommends shortening wholesale funding by USD 1.2 billion over the next two stages" is a correct mathematical answer and a defensible explanation in roughly zero ALCO venues. The OCC's Interest Rate Risk Comptroller's Handbook expects directors to challenge assumptions and confirm that the methodology behind any material strategic recommendation is documented at a level they can understand (OCC 2020). Without a translation layer between the solver and the ALCO seat, ALCO defaults to qualitative heuristic decision-making: peer benchmarking, prior-period continuity, conservative carry-over.

The strategic playbook

The realistic posture is to stop asking the algorithm to invent the strategy. ALCO defines the macro strategy (raise USD 1 billion in term funding over six months to fund loan originations, or extend the duration of the securities book by twelve months to lock in current yields). The algorithm is given a discrete, time-bounded menu of pre-approved management actions and is asked to find the best feasible tactical mix from that menu, against the constraint stack, across a multi-scenario evaluation matrix.

The action library is the institution's playbook: specific swap overlays in named tenors with named notional bands, funding-mix shifts (terming out wholesale, brokered-CD issuance windows, FHLB advance ladders), FTP recalibrations, hurdle-rate moves on origination, HQLA composition adjustments. Each action is operationally feasible, regulatorily compliant at execution, and pre-approved through the relevant policy committees. The 2016 interagency guidance on FTP (FRB, OCC, FDIC 2016, SR 16-3) frames FTP as the enforcement layer for the contingency funding plan; the same logic applies to the production playbook.

Lubinska 2020 documents the structure at scale. The four worked case studies in chapter 4 converge to roughly the same funding-mix proportions across scenarios because the binding constraints, typically Survival Horizon and the NII-sensitivity limit, dominate the optimum. The funding mix is conditioned on the constraint stack rather than on point estimates of cost. That convergence is the operational signature of a working playbook. The Financial Risk Academy ALM Modeling and Balance Sheet Optimization course implements the demand-curve mechanism that makes pricing endogenous and the Markovian regime-switching scenario generator that runs the action library across BAU, deterministic regulatory shocks, and stochastic rate paths.

The output is implementable by construction because the menu was pre-screened, defensible because each action carries its own committee sign-off, and validatable because the action library is the unit of governance rather than the dual variables on the constraints. Inside that frame, a handful of sub-problems are tight enough to formalize at full optimization rigor. Non-maturity deposit replicating portfolios are the canonical case; I worked the field literature in Replicating Portfolios in Banking: A Literature Review on Asset Liability Management (March 2025) and made the case that optimization-based ladders dominate static maturity assumptions for NMD replication. HQLA composition under LCR constraints, FTP curve calibration against a defined funding plan, and capital allocation via Euler-allocated economic capital have the same character. The cross-problem trade-offs belong in the playbook layer, where ALCO names the position the institution chooses to own and sizes the action library against it.

Agentic AI as the orchestration layer

Here is where the picture changes. The strategic playbook has been the right answer for at least the last decade. What has held it back from running as the institution's decision engine is not the math. It is the cost of stitching together four calculation platforms that were never designed to talk to each other, and the cost of running the playbook across a wide-enough scenario matrix on a cadence that keeps up with the balance sheet. Building a single integrated surface across the four platforms has historically required a multi-year systems-integration program that no bank wants to fund.

Agentic AI changes the economics of that integration. Not by replacing the quant models, which remain what they were, but by becoming the orchestration layer that sits above them. The way I think about it: the mathematical models are plugins, and the agent is the operating system that calls them. A few pieces of this are now realistic in a way they were not two years ago.

The first is data harmonization. LLM-based parsers can read the schema of each silo's platform on the fly and produce a unified view of the balance sheet without anyone hand-coding ETL pipelines. The data lineage problem that has bedeviled BCBS 239 implementations for a decade is, in 2026, a problem an LLM can do real work on. The harmonization is not perfect, but the comparison is no longer between an integrated platform and a fragmented one; it is between a fragmented platform and one the agent can read across.

The second is tool-calling. An agent given the right wrappers can invoke the cash-flow engine, the LCR runoff model, the RWA tracker, and the FTP curve constructor on demand and pass results between them in a single thread. The orchestration logic that previously required a quant analyst's Python wrapper and a vendor update to three platforms can now sit in a system prompt the ALCO chair could conceivably read.

The third is semantic guardrails. ALCO's risk-appetite policy and the regulatory constraint corridors can be encoded as natural-language constraints checked against every proposed strategy. The agent does not need to compute the dual on the NSFR constraint to know that a recommendation breaching the contingent funding plan is out of policy; the policy text is part of the agent's context.

The fourth is the iteration loop. Give the agent an ALCO-set goal ("raise USD 1 billion in term funding at the lowest all-in cost given current LCR headroom"), the pre-approved strategic menu, and the constraint stack, and the agent can screen the option space, run each candidate against the multi-scenario matrix, and rank by risk-adjusted return against named ALCO preferences. The output is not a dual variable. It is a ranked shortlist of named actions, each with its scenario performance, its constraint utilization, and the reasoning that produced it.

None of this dissolves the four structural impediments. Behavioral inputs are still uncertain; the constraint corner still rotates; the four silos still own four platforms; ALCO still has to take a position the model cannot take for it. What changes is that the cost of running the playbook against a realistic scenario matrix at the cadence the balance sheet needs has dropped by orders of magnitude. The investment shifts from writing raw math code to what I would call context engineering: encoding the action library, the risk-appetite policy, and the scenario matrix in a form the agent can use. The mathematical capability has been mature for decades. The orchestration capability is what was missing.

Close

The toolkit is mature. The silos are the bottleneck. The agentic orchestration layer is what finally makes the strategic playbook tractable across the four platforms at the cadence the balance sheet needs. The investment the institution has to make is not in mathematical capability, which is already there. It is in defining the rules, the goals, and the action library across the four areas of subject-matter expertise in a form the agent can read, iterate against, and bring back to ALCO as a ranked shortlist. The agent does not solve the optimization. It connects the bank's experts to solve it together. The question worth taking to the next ALCO meeting is which sub-problems the institution is prepared to formalize at full optimization rigor, and which trade-offs belong in the rule-based, scenario-driven, action-library layer where ALCO names the position the institution chooses to own and the agentic orchestration runs.

Disclaimer: The views and opinions expressed in this article are my own and do not reflect the views of my employer or any organization I am affiliated with. This content was written with the assistance of AI tools and is provided for informational and educational purposes only. It is not intended as financial, legal, or regulatory advice. Readers should consult with qualified professionals before making any decisions based on this content.

References

Adam, Alexandre. 2008. Handbook of Asset and Liability Management: From Models to Optimal Return Strategies. Chichester: Wiley.

Basel Committee on Banking Supervision. 2013. Basel III: The Liquidity Coverage Ratio and Liquidity Risk Monitoring Tools. Basel: Bank for International Settlements, January. https://www.bis.org/publ/bcbs238.htm

Basel Committee on Banking Supervision. 2014. Basel III: The Net Stable Funding Ratio. Basel: Bank for International Settlements, October. https://www.bis.org/bcbs/publ/d295.htm

Basel Committee on Banking Supervision. 2016. Interest Rate Risk in the Banking Book. BCBS 368. Basel: Bank for International Settlements, April (consolidated into SRP31, effective 1 January 2026). https://www.bis.org/bcbs/publ/d368.htm

Chen, Chih Lung. 2025a. "Replicating Portfolios in Banking: A Literature Review on Asset Liability Management." LinkedIn, March.

Chen, Chih Lung. 2026. "Stable, Structural, Core: One Word, Several Conditions." LinkedIn, May 11. https://www.linkedin.com/pulse/stable-structural-core-one-word-several-conditions-chih-chen-kk1yc/

Choudhry, Moorad. 2018. The Moorad Choudhry Anthology: Past, Present and Future Principles of Banking and Finance. Chichester: Wiley.

Coulier, Lara, Cosimo Pancaro, Edoardo Pancotto, and Alessio Reghezza. 2025. "Banking on Assumptions? How Banks Model Deposit Maturities." European Central Bank Working Paper No. 3140.

Dermine, Jean. 2003. Asset and Liability Management: The Banker's Guide to Value Creation and Risk Control. Harlow: Pearson Education / Financial Times Prentice Hall.

Drechsler, Itamar, Alexi Savov, and Philipp Schnabl. 2021. "Banking on Deposits: Maturity Transformation without Interest Rate Risk." Journal of Finance 76 (3): 1091-1143.

Drechsler, Itamar, Alexi Savov, Philipp Schnabl, and Olivier Wang. 2026. "Deposit Franchise Runs." Journal of Finance, forthcoming.

European Banking Authority. 2022. Guidelines on the Management of Interest Rate Risk Arising from Non-Trading Book Activities (IRRBB) and on the Treatment of Credit Spread Risk Arising from Non-Trading Book Activities (CSRBB). EBA/GL/2022/14, October.

Federal Reserve Board, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation. 2016. Interagency Guidance on Funds Transfer Pricing Related to Funding and Contingent Liquidity Risks. SR 16-3, March 1.

Federal Reserve Board, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation. 2026. Revised Guidance on Model Risk Management. SR 26-2, April 17.

Financial Risk Academy. 2026. Building an ALM and Balance Sheet Optimization Model. https://financial-risk-academy.teachable.com/courses/alm-asset-liability-management-training

Lubinska, Beata. 2020. Asset Liability Management Optimisation: A Practitioner's Guide to Balance Sheet Management and Remodelling. Chichester: Wiley.

Office of the Comptroller of the Currency. 2020. Interest Rate Risk. Comptroller's Handbook, June.

Sodhi, ManMohan S. 2005. "LP Modeling for Asset-Liability Management: A Survey of Choices and Simplifications." Operations Research 53 (2): 181-196.

Van Deventer, Donald R., Kenji Imai, and Mark Mesler. 2013. Advanced Financial Risk Management: Tools and Techniques for Integrated Credit Risk and Interest Rate Risk Management, 2nd ed. Singapore: Wiley.

About The Author

Chih Chen

is a bank treasury and risk leader with over 20 years of experience in Asset Liability Management and IRRBB. Currently SVP, Treasury at East West Bank, he focuses on translating quantitative insights into earnings stability and risk strategy.

He has previously held leadership roles at BNP Paribas (Bank of the West) and MSCI Inc., and is a published voice on deposit behavior and ALM.