Silicon Valley Bank did not fail because it could not measure its interest rate risk. It measured it, breached its own limits, and then changed the assumptions until the numbers behaved. The risk stayed.
March 2023 ended the era in which interest rate risk in the banking book (IRRBB) could be treated as a second-tier risk: real, but rarely fatal. SVB, then the 16th largest US bank, failed in substantial part because of textbook IRRBB, unmanaged. The numbers, assembled from the Federal Reserve’s Barr report and public filings, map onto the discipline’s core concepts almost too neatly. It is a story in three acts.
During the deposit boom, SVB’s deposits tripled in three years to roughly $190bn. Not just any deposits: large, uninsured, rate-sensitive corporate balances, concentrated in a single ecosystem of venture-backed technology firms. In the vocabulary of deposit modeling, these were the least “core” non-maturity deposits (NMDs) imaginable: contractually withdrawable overnight, behaviourally correlated with each other, and belonging to clients who watch rates for a living.
Management invested the inflow in long-duration, fixed-rate securities: a portfolio of roughly $120bn, about $91bn of it classified as held-to-maturity (HTM), yielding around 1.6% with duration in the region of six years. The result was an enormous long-duration position, equity duration strongly positive, entirely exposed to the parallel-up rate scenario. On the economic value lens, the bank was one sustained rate rise away from serious trouble. On the earnings lens, it looked comfortable.
That divergence between the two lenses, economic value of equity (EVE) and net interest income (NII), is the single most important piece of intuition in IRRBB. The two measures routinely give opposite signals on the same balance sheet. SVB is the cleanest real-world illustration on record, because when the signals diverged, the bank chose to look only at the comfortable one.
In 2022, the Federal Reserve raised rates by more than 400 basis points. SVB’s internal EVE metrics breached the bank’s own limits. What followed was not a repositioning of the balance sheet. It was a repositioning of the measurement.
The response sequence, as the record shows it: hedges that protected against rising rates were unwound to support current earnings. The deposit duration assumptions in the EVE model were lengthened, which moved the measure back inside limits without changing a single position. And management attention stayed on NII, which rising rates initially flattered, because new lending and reinvestment reprice upward faster than sticky deposit rates rise.
SVB did not fail to measure its risk. It measured its risk, disliked the answer, and re-measured until the answer went away.
None of this was invisible. Every step was seen and noted by supervisors as outstanding findings. Escalation simply did not keep pace with the position. The detection worked; the consequence did not, a distinction that now shapes the regulatory debate in every jurisdiction.
By end-2022, unrealized losses on the securities portfolio, roughly $15bn on the HTM book alone, approximately equaled the bank’s total equity. That fact was absent from regulatory capital, because of the AOCI opt-out and HTM’s amortized-cost treatment. It was not absent from the accounts. It sat in the footnotes, published, for anyone who cared to do the arithmetic.
On 8 March the bank sold its roughly $21bn available-for-sale book at a $1.8bn realized loss and announced a capital raise. The signal converted a rate-risk problem into a solvency question, and the depositors, uninsured, concentrated and connected, answered it. $42bn of deposits left the next day, with $100bn more queued behind. The bank failed on 10 March. Two years of interest rate risk crystallized through the liquidity channel in roughly 48 hours, at social-media speed.
This is the interaction that makes IRRBB dangerous. Rate risk rarely kills alone. It waits inside amortized-cost accounting until a liquidity event forces the sale that makes the losses real, and by then the depositors are already moving.
The regulatory post-mortems each drew their own conclusions, and there is a live argument about which framework would have caught it: an EU-style outlier test would have flagged the position loudly, and the UK’s supervisor-computed capital charge would have priced it years early. That comparative question deserves its own treatment.
But the deeper lesson sits below the frameworks, in the models. Every IRRBB number a bank produces, every ΔEVE, every duration of equity, every limit utilization on every ALCO pack, rests on behavioral assumptions about deposits: how much is stable, how long it stays, how fast its rate reprices. Lengthen the assumed deposit duration and measured risk falls. No position changes. No hedge is bought. The risk is exactly what it was; only the number has moved.
That is why all four major regimes emerged from 2023 sharing one operational conviction: deposit behavior models calibrated on pre-digital, pre-2022 data are the soft underbelly of every IRRBB number, and NMD assumptions are now the common thread of supervisory programs everywhere.
SVB’s epitaph belongs in every model governance paper: every IRRBB number rests on deposit assumptions, and assumptions can be managed in both senses of the word.
For the full picture, including the Basel BCBS 368 framework, the deposit-modeling debates behind the SVB failure, and how the EU, UK, and US regimes each address it differently, read the complete guide "IRRBB: A Complete Reference, From Zero to Expert."
To see how Mirai Regulatory Reporting automates IRRBB measurement, EVE and NII stress testing, and regulatory reporting across these frameworks, learn more here.