Bank securities marks under zero-trust — "held to par" is a choice, not a price

Web-verified 2026-06-09. Structured + edges + sources: macro-bank-htm-marks.json. The asset-side mirror of the AI paper-marks proof (reflexive_marks). Honest update: aggregate unrealized losses fell through 2025 — the point is structural fragility + opacity, not a current acute hole.

The numbers — full FDIC QBP series — STRONG where FDIC-stated

Unrealized losses on HTM + AFS securities (the hole tracks long rates and reopens whenever they rise):

Confidence: Q4-2022 / Q4-2024 / all 2025–26 quarters are FDIC-stated; the Q3-2022 (~$690B) peak and Q3-2023 ($684B) / Q4-2023 ($478B) are widely-reported approximations; 2019–21 is qualitative. Exact per-quarter figures live in each QBP PDF.

• The 2025 decline was driven by falling long rates, not accounting; the Q1-2026 reversal (losses up as the 30-yr mortgage rate rose) proves it's rate-dependent and not gone.
• HTM = carried at amortized cost (value doesn't move with the market unless sold); AFS = marked-to-market through AOCI (and, for big banks, regulatory capital).
• Profits hit records alongside this (Q1-2026 net income $80.5B, ROA 1.26%), and the problem-bank count fell to 54 — but the FDIC stopped publishing problem-bank asset totals in 2025, so the size of concentrated exposure is no longer disclosed.
• The BTFP — the Fed facility that let banks borrow against underwater bonds at par — expired March 2024, removing the explicit 2023 backstop.

The zero-trust read

• HTM is a choice. Whether a loss appears depends on a classification the bank picks. Move a bond to HTM and the mark-to-market loss leaves reported equity. "Book equity" and HTM "held-to-par" are self-reported solvency, not prices — the same critique applied to government statistics and to AI fair-value marks, now on the asset side of banks.
• Unrealized until forced. The loss stays invisible only while the bank isn't forced to sell. A deposit run forces sales, realizes the loss, and can taint the whole HTM book — the SVB 2023 mechanism: solvent at cost, insolvent at market, dead in a weekend.
• Opacity. The FDIC stopped publishing problem-bank asset totals in 2025 (macro-fdic) — the aggregate is visible, the concentration is not.
• Symmetry with the AI marks. AI funders hold private stakes up at self-set marks and book gains (reflexive_marks M2a); banks hold bonds at cost and avoid losses. Both are "carried at a chosen value, not a market price, until an external event forces realization" — an IPO below the mark for the AI side, a deposit run for the bank side.

Exposing what the withdrawn aggregate hides

The FDIC dropped the problem-bank asset total, but that's cosmetic — the facts are still recoverable:

Reconstruct it directly (the number wasn't deleted, just the summary):

• Per-bank Call Reports (FFIEC 031/041, Schedule RC-B) report HTM amortized cost and fair value, AFS fair value, and pledged securities — so each bank's unrealized HTM loss = fair value − amortized cost is computable per institution, every quarter.
• FDIC BankFind Suite API (banks.data.fdic.gov, no registration) serves institution-level financials incl. securities, equity, and uninsured deposits. This repo already does it: models/graph/bank_exposure.py pulls the API and computes per-bank HTM-loss/equity (e.g., BofA HTM −$81B ≈ −34% of equity).
• Uninsured-deposit share (call report) — the run-risk metric (SVB was ~94%); high uninsured % + big HTM loss = the SVB setup.

Funding-stress proxies (FDIC-aggregate-free):

• FHLB advances — the "lender of next-to-last-resort": system advances $742.8B (Jun 2025) vs $736.7B (YE 2024), peaked Mar 2023; growth at higher-risk banks is a tell.
• Fed discount window / H.4.1 — aggregate primary credit weekly; spikes flag stress even though names lag ~2 years.
• BTFP run-off (expired Mar 2024) mapped who was underwater; brokered/hot-money deposits show funding fragility.

Market proxies (front-run the accounting): KBW regional-bank index (KRE), single-name drawdowns, bank CDS spreads, options skew, rating watch lists — and the FDIC still publishes the problem-bank count and the failed-bank list.

Opacity of the aggregate ≠ opacity of the facts. The per-bank RC-B + FDIC API rebuild the exact number that was withdrawn (bank_exposure.py does), and FHLB advances / discount-window spikes / CDS / uninsured-deposit share give independent stress signals.

Posture

Read "book equity" and HTM "held-to-par" as the most solvent admissible presentation, and stress-test at market (AFS + HTM marked) and under forced-sale — not at cost.

Where this sits in the larger pattern (updated 2026-06-11)

HTM amortized cost is the first of five self-marked numbers: bank securities (here), AI fair-value marks (fin-google-amazon-anthropic-meta), private-credit NAVs (macro-private-credit-marks), insurance captive marks (spec-insurance-bermuda), and AI-compute depreciation (fin-ai-depreciation-debttrap). The self_marked_value proof (U1–U4) shows the four balance-sheet marks are one defect whose gaps correlate under a common factor, so "diversified across asset classes" is false; the cross-sectional analysis measures that factor at ~91% of credit-spread variance. The bank-HTM hole is the original instance (SVB, 2023) — and the same rate move that re-opens it is the common factor that simultaneously prices the other four.

← Research index · structured data: macro-bank-htm-marks.json · macro-bank-htm-marks.md