The one series that saw the panic is published two years late
In the quarter Silicon Valley Bank failed, banks drew more from the Fed's discount window than in the whole of the rest of the disclosed record put together. By law, nobody outside the Fed could see it for two years. That is not a scandal. It is the price of the window working at all, and it puts a hard ceiling on what any early-warning system can do.
Read the units before the numbers.
Every figure below is originations: the sum of loans made during the quarter. It is not the balance outstanding. Discount-window loans are typically overnight, and they are rolled, so the same dollar borrowed and repaid each day is counted again each day. The flow is therefore far larger than the stock, and $3.14tn of quarterly originations does not mean $3.14tn was owed to the Fed at any instant. Anyone quoting this series as an outstanding balance is misreading it.
With that said, here is what the Federal Reserve’s discount-window disclosures show for 2023Q1, the quarter in which Silicon Valley Bank failed ($209.0bn in assets, 10 March 2023) and Signature Bank failed ($110.4bn in assets, 12 March 2023).
Primary credit originations were $3.14tn, across 3,519 loans to 701 distinct borrowers. The quarter before, 2022Q4, the figure was $172.6bn. That is a jump of 18.2 times in three months.
Two comparisons give it scale. The first month of the pandemic, 2020Q1, produced $89.3bn of originations: 2023Q1 is 35 times that. The second is the one that stops you. Across the whole disclosed record, 56 quarters from 2010Q3 to 2024Q2, primary-credit originations total $4.18tn. A single quarter, 2023Q1, accounts for 75.1% of it. Three quarters of everything the discount window has done in fourteen disclosed years happened in those three months.
And you were not allowed to know
The discount window is the clearest distress signal in banking. A bank does not borrow from the lender of last resort because the terms are good. It borrows because the alternatives have closed. When 701 institutions draw $3.14tn in a quarter, something is being said out loud, in the only language a balance sheet has.
Nobody outside the Federal Reserve could hear it. Loan-level discount-window disclosures are published on a roughly two-year statutory lag, mandated by section 1103 of the Dodd-Frank Act (12 U.S.C. 248(s)). The 2023Q1 data did not become public in 2023Q1. It became public around two years later. The most recent quarter FinObservatory can hold is 2024Q2, and that is not a gap in our collection: it is the whole of what exists.
So the sequence runs: the banks borrow, the Fed sees it, the panic resolves or it does not, and the public gets the numbers when they are history. In the window that mattered, from the first withdrawal to the last failure, this series was empty.
The lag is the point, not the failure
The obvious reaction is that the lag is a scandal, and it is worth resisting, because the lag is load-bearing. Borrowing from the discount window carries stigma: a bank known to be at the window is a bank whose depositors are told, publicly, that it could not fund itself anywhere else. If that were disclosed in real time, the window would stop working, because the banks who most need it would not dare use it, and the facility that exists to stop runs would start them. Congress set a lag long enough that the disclosure cannot trigger the run it is describing.
That is a defensible trade, and it is not a free one. The price is exactly this: the best real-time indicator of bank distress in the United States is, by construction, not available in real time. Not to us, not to depositors, not to any model.
What this means for every early-warning system, including ours
FinObservatory’s own bank-failure nowcast had Silicon Valley Bank at the 43.8th percentile of vulnerability on its last call report before it failed: below the median, against an average of the 77th percentile for banks that actually failed. We publish that miss rather than bury it. This brief is part of the reason for it.
A model fed quarterly call reports is looking at accounting ratios filed weeks after the quarter ends. The series that would have screamed, this one, is embargoed for two years. So the sharpest available signal is structurally unavailable to any monitor, public or private, at the only moment it would be worth having. That is not a modelling problem, and no amount of model class, feature engineering or cross-validation touches it. The information is not late in our pipeline. It is late in the law.
It is worth being precise about the claim. This does not say the 2023 failures were unforeseeable: the interest-rate risk in SVB’s securities book was visible to anyone who read the filings, and the concentration of its uninsured deposits was public. It says something narrower and more uncomfortable. The one series that records, directly and without inference, which banks could no longer fund themselves, is a series that arrives after the story is over. When you read that anyone should have seen it coming, ask which data they mean, and check when it was published.
Every figure above is queried from the Federal Reserve’s discount-window disclosures when this page is built, aggregates only, and reported as originations rather than balances. The series and its statutory basis are in the data catalog.