Third Derivative Research · the answer so far
Why US equity prices rose and where the money came from. Every claim carries a grade; the link between flows and prices cannot be measured on free data.
Why prices rose: mostly because earnings rose, and most of that earnings growth is real profit, not buybacks. Where the money came from: the line the Federal Reserve's accounts show as equity "issuance" is mostly ETF share creation, and US operating companies retired equity, on net, over 2024–26. The join between the two halves, how much a dollar of buying moves a price, cannot be measured on free data; four attempts failed, each for a stated reason (§06). Government bonds, currencies and commodities are covered too: government bonds as first passes, currencies and commodities as a first map and four tests. We report what was measured (how yields split, who held the new supply) and what moved together. None of it shows cause.
Prices rose mostly because earnings rose, and most of that earnings growth is real profit, not buybacks shrinking the share count. The multiple held up because the earnings yield (earnings divided by price) ended the window close to where it began while real yields rose about a third of a point. Inside the window the earnings yield fell about 0.6 point in 2024 and regained most of that in the first half of 2026. Calling that a “compressing risk premium” adds less than it seems. The premium is defined as the earnings yield minus the real yield, so most of its fall is the same rate rise, restated. In 2024 the multiple's expansion came as drift, not on scheduled news.
On the money side, the main finding is a measurement result, not an economic one. The line everyone reads as "equity issuance" in the Federal Reserve's accounts is mostly ETF share creation (an ETF, or exchange-traded fund, is a fund whose own shares trade on an exchange). Strip that wrapper out and US operating companies (non-financial firms) retired equity over 2024–26. The money arriving met a shrinking float of company stock (the float is the stock available to trade).
Across the channels the goal names, the verdict is mostly negative. No measure shows collateral re-use rising before the price rise. The one dated step in the banks' collateral stock came in the first half of 2026, after most of the rise, and earlier years are not measured. In 2024, new shadow money was about a tenth of households' purchases of equities, after netting out cash that simply moved from the Fed to dealers. Over the whole window it stays in the picture as a ceiling. The AI build-out is funded mostly from operating cash flow. The real financing innovation is off balance sheet, in lease commitments. The AI debt that does exist sits in a corporate bond market held mainly by long money (insurers, pensions, bond funds and overseas holders). The AI names are not a separate line in Z.1 (the Fed's quarterly financial accounts), and free data count only two holder groups for them. A name search finds money-market funds holding about $3.4bn under AI names, mostly as repo collateral (repo is short-term borrowing against securities); that figure is a floor. The IPO boom is between half and two-thirds a single deal, depending on which total is used (§04).
Government bonds are now covered, as first passes. Since end-2023 the 10-year yield rose 113bp in the US, 156bp in Germany and 243bp in Japan (a basis point, bp, is a hundredth of a percentage point). In the US almost all of it was real yields (yields after inflation, read from inflation-protected bonds). In Japan inflation compensation did between a third and two-fifths. In Germany the rise was real in 2024–25: one inflation-linked bond's breakeven fell, though part of that may be a liquidity premium (the breakeven is the inflation rate implied by the gap between ordinary and inflation-linked yields). In 2026, German market breakevens did about half or more. The Japanese public's rate risk in Japanese government bonds (JGBs) grew 11–25%, depending on the measure and the start date. In all three markets the central bank shrank its holdings while others absorbed new supply. Which investors did so differs by market. Whether that supply moved yields is not measured (§05).
What we cannot say is how much of the price move the money caused. Four designs to measure that have failed on free data, each for a different, stated reason (§06).
This page is the project's authoritative overview (ruling of 14 October 2026). The evidence behind each claim, with its grade, the findings note that owns it and its caveats, is in 2026-09-17-ANSWER-What-Is-Driving-Asset-Prices.md and the notes it names. Where that note and this page differ, one of them is wrong and is corrected. Claims that failed are kept, not deleted, in CORRECTIONS.md; most sections end with the correction numbers (C-NNN) and notes behind it. The page's source is docs/index.html in the public repository, and the project's checks scan it against the corrections ban list on every run.
The grading rule is the principal's ruling of 22 September. “Nothing here is certain if not clearly mathematically proven, thus everything needs to be looked at as potential hypothesis that needs testing.” So every claim carries one of four grades, and the default is the weakest. Identity: true by construction, and therefore not evidence; a sum that cannot fail proves nothing. Measured: observed on a stated universe with a stated denominator. Bounded: a ceiling or a floor, never quoted as an estimate. Hypothesis: untested, or tested and survived; survival is not proof. The ledger in §08 carries these grades.
AI models wrote this page for the principal, the person who directs the project, sets its question and rules on what counts as evidence: Claude (Anthropic) and Grok (xAI) as peers, with Gemini (Google) consulted for reading and online research. It is that person's personal opinion and has nothing to do with their employer. It is not advice, and errors are expected. Full disclaimer.
Sources and corrections: C-151 (the earnings yield's path inside the window); C-102 (the premium is the rate rise restated); C-118 and C-152 (shadow money, collateral re-use); C-139 and C-138 (who holds the AI debt; the money-fund floor); C-128 (who absorbed bond supply). The rates work is first passes reviewed from outside; the 8 October additions (Germany 2024–25, Japan's duration supply) were reviewed on 11 October, and the German split stood. Register: CORRECTIONS.md.
Two decompositions, each reproduced from source data: where the value sits, and what the earnings part is made of.
Our premium measure has no growth term, so we checked it against one that does. Damodaran's forward-looking, growth-adjusted equity risk premium fell 4.60% → 4.20% over the identical months. Ours fell 49 basis points, his 40 and a survey measure 19. They agree for a shared structural reason, not independently: in all three a bond leg rose faster than the equity leg (+34bp, +57bp, +35bp), and in the survey, equity expectations actually rose. Our measure then gave back about 17bp in the first half of 2026 as earnings caught up; that part is measured.
What did not resolve: whether the compression was cheaper risk or better growth expectations. A survey measure does not settle it: the survey premium fell on its bond leg while equity expectations rose. Netting growth out with Damodaran’s model leaves 36–60% of our 49bp unexplained by growth, depending on which of two defensible rate bases is used. The two ends come from the two rate bases. The 36% end is not a floor. On the growth sensitivity seen in a 2015–19 control window, growth could account for all of the 49bp. That rests on two windows and is not a finding.
Sources and corrections: C-119 (the Fed's reclassifications); C-080 (the two misreadings of the 7%); C-102 (the rate/premium split, and 71% of the premium leg); C-101 (the buyback ceiling); C-082 (the survey's risk–growth split is withdrawn); C-104 (36–60%); C-153 (36% is not a floor). Register: CORRECTIONS.md.
The amount of equity is not fixed. What grew was the fund layer on top of company stock, not company stock itself.
The Federal Reserve's "corporate equities" instrument counts ETF shares as equity issuance. ETFs are a wrapper layer above company stock, and some of that layer is backed by bonds. Classified properly, 2024:Q1–2026:Q2 reads: ETF share creation +$3,603.8bn and operating companies (non-financial corporates) −$540.7bn. The ETF figure matches the Investment Company Institute’s own census to under $0.1bn in both years. The Fed’s ETF table appears to be built on that census, so the match shows the figure was carried over correctly. It does not show that a second source agrees. “Issuance net of the wrapper”, −$360.7bn, is a derived blend that also nets in foreign issuers, so it understates the operating-company retirement by about a third. The Fed’s issuer lines summing to the total is how its table is built, not a check. The bars below are real, but they measure the growth of the fund layer.
Gross retirement was mostly buybacks. On the Fed's Enhanced Financial Accounts for 2024:Q1–2026:Q2, operating companies retired $2,979bn (65% repurchases, 35% M&A) and issued $2,427bn, leaving net retirement of $552bn. This is a different series from the −$540.7bn above, over the same window: that figure is the Z.1 net flow for non-financial corporates, while this one is the net of the gross flows in the enhanced accounts. The −$360.7bn blend is neither. The issuance leg includes private placements while repurchases are public firms only, so the two are not on one basis. No single channel "accounts for" a net, M&A included.
Over the decade the Fed's figures show buybacks of $6.40trn against issuance of $6.62trn, but the two cannot be set against each other. The Fed’s own methodology counts buybacks by public firms only, while its issuance series blends public and private, and only 18% of that issuance was identifiable public offerings. Against those, buybacks were 5.3 times larger. The near-equality comes from netting in private placements that never touch a listed share count. That is also why the buyback test in §02 used company-level filings instead.
Still unreconciled: two Fed series for non-financial net issuance differ by an offset that varies from year to year (+$45.4bn against +$51.3bn for 2026:Q1 on current vintages; the cause is not identified). They must not be mixed in one table.
Sources and corrections: C-081 and C-083 (ETF shares counted as issuance); C-148 (the ICI match is a carry-over); C-107 (the −$360.7bn blend); C-106 (issuer lines summing to the total); C-108 (the two legs are not on one basis; the decade comparison is withdrawn); C-133 (no single channel accounts for the net); C-130 (the unreconciled series). Register: CORRECTIONS.md.
The goal names each channel. Most verdicts are negative, and each comes with its reason.
| Channel | Verdict | Over 2024–26 |
|---|---|---|
| Passive bid / ETF creation | Dominant | The largest part of net new equity: the wrapper is the issuance (§03). ETFs bought about $2.46trn while mutual funds sold about $2.01trn, an 82% offset, consistent with rotation. Whether it was the same money switching wrapper is not measurable on free data. |
| AI build-out funding | Mostly self-funded | Three of the five large spenders generate more operating cash than their capital spending; the financing that matters is off balance sheet.
|
| IPO boom | Measured on eight deals; buyers unknown | About $112bn of new company money in 2026, to 12 September, across the eight 2026-priced operating-company deals among the ten prospectuses we read. One deal, SpaceX ($75bn), is two-thirds of that $112bn, and about half of Renaissance Capital's total of $146bn (a vendor figure; its universe is not stated). Who was allocated the shares is not recoverable on free data; named cornerstone investors, committed before the listing, appear in 3 of 10 deals. |
| Direct money creation | Weak, not a ceiling | A small positive association with equity buying. The deposit stock is 6.3× the purchases, and ETF creation is often in-kind, so deposits bound nothing. |
| Shadow banking / shadow money | Not in 2024; a ceiling over the whole window | In 2024, a handoff, not fresh money: the increase in wholesale repo came out of the Fed's reverse-repo facility, where mainly money funds lend cash to the Fed. Over the whole window it is not ruled out: raw growth is the same order of magnitude as the purchases, and even net of the handoff about a quarter of them. No share of purchases can be attributed to nonbank money without inventing one. The wholesale share of nonbank funding has also stopped rising: 23.07% of wholesale funding plus M2 (the broad money supply) in March 2026 and 22.96% in June. That is a plateau: the 0.11pp gap would vanish with a 0.6% (about $145bn) revision to June M2. |
| The collateral channel | Not the source | No measure here shows collateral re-use rising before the price rise. The large banks' collateral stock stepped up in the first half of 2026, after most of the price rise. The share of it they re-pledged edged up in the same half-year, from 81.7% to 83.2%; earlier shares are not measured. |
On these measures, the money that bought equities was mostly already inside the system: the deposit stock, and offsetting flows between fund types. It was mostly not newly created through shadow or collateral channels; shadow money stays in the picture as a ceiling. The deposit stock is 6.3 times the purchases, so it bounds nothing. The largest capital-spending programme in the market was funded mainly from profits and leases. That does not make those channels unimportant in general. It says nothing measured shows them growing before the price rise, or on the scale it would have needed. Whether the lease stack is the next source of fragility is a different question; its tests are dated in §10 (the third-quarter filings in November, Oracle's guarantee in December).
Sources and corrections: C-111 (same money switching wrapper is not measurable); C-134 (the firm-level cash-flow counts); C-116 (the lease-to-guarantee ratio, like for like); C-139 (holders of the AI debt; the Beignet share); C-138 (the money-fund name search); C-115 and C-120 (bank exposure); C-089 (deposits bound nothing); C-085 (the 2024 reverse-repo handoff); C-118 (shadow money as a ceiling); C-130 (the wholesale-share plateau); C-152 (collateral re-use). Register: CORRECTIONS.md.
Rates are in the goal's scope. Three first passes cover US Treasuries, Japanese government bonds and euro-area bonds, from end-2023 to late September 2026.
| Market | 10-year yield | What moved it | Who took the supply |
|---|---|---|---|
| US Treasuries | 3.88% → 5.01%, +113bp to 18 Sep | Real yields: TIPS (Treasury inflation-protected securities) +96bp, breakeven +17bp. Models and the survey agree that expected rates for the next four years were lower at the August 2026 survey than at the first, in February 2024: two statistical models by −60bp and −24bp, and the forecasters' survey by −19bp. That fall came between the early-2024 and early-2026 surveys, and half to two-thirds of it reversed after that. Most of the 113bp came after 2 March 2026. Whether the rise beyond that fall in expected rates is term premium (the extra yield for holding a long bond instead of rolling short ones) turns on years five to ten, where the models and the survey part ways. | The public took more rate risk, mostly from Treasury issuance; registered funds are not established as a first-order holder.
|
| Japan | 0.65% → 3.07%, +243bp | Breakeven inflation rose 82–97bp, depending on which inflation-indexed issue is used: a third to two-fifths of the rise. Since summer 2024, the Bank of Japan (BoJ) reads expected rates and the term premium as contributing about equally, with about 25bp from its purchase cuts. | The BoJ's holdings fell by a net ¥63trn as bonds matured, not through sales. Banks bought ¥55trn, public pensions ¥32trn and overseas investors ¥24trn (2024:Q1–2026:Q2, Flow of Funds). These purchases are not a split of the BoJ's fall: they also took up new issuance. At constant yields the public took 16–25% more fixed-coupon JGB rate risk (31 Aug 2026 against Oct 2024 and end-2023) on the method used for the US, and 11–18% on an unpriced measure. Since end-2023 that was ¥138trn of 10-year equivalents, about 1.7 times the ¥83trn rise in total supply. Those are upper-end figures on the end-2023 curve; the size depends on the curve (11–25% across measures). |
| Euro area (Bund) | 2.06% → 3.62%, +156bp | Mostly real over the window. In 2024–25 the breakeven on the 2033 linker (an inflation-linked Bund, the German government bond) fell 21bp while the nominal yield at its maturity rose 66bp (Bundesbank daily files). Part of that may be a liquidity premium, since issuance stopped in 2024. But breakevens on inflation-linked Bunds rose 34–49bp in 2026 against the Bund's 68bp, so inflation compensation did about half or more of this year's rise. The ECB reads the rise since late 2024 as largely higher real term premia. | Holdings of the Eurosystem (the ECB and the euro-area national central banks) fell €751bn; banks bought €725bn and the rest of the world €830bn (face value, 2023:Q4–2026:Q2). These purchases are not a split of the Eurosystem's fall: they also took up new issuance. The Eurosystem's share fell from 31% to 21%. |
Supply is shared; who took it differs by market. In all three markets the central bank's holdings were lower in September 2026 than at end-2023 while the government issued. The absorbers differ: banks and public pensions in Japan, banks and foreign investors in the euro area, and in the US a household line that is likely to be largely offshore hedge funds. Holders lost against cash in all three. A constant-maturity 10-year returned about +3.5% in the US against +12.9% for bills, −4.7% in Germany against +6.9% for one-year paper, and −13.9% in Japan against +1.6%. None of this shows that the supply moved yields. The term premium is not measured on free data in any of the three, and the charter forbids reading flows as cause.
In the swap market, Treasuries yield more than swaps at every maturity measured, 2 to 30 years. On the public tape of US-reported swap trades (DTCC), the 10-year Treasury yielded 39bp more than the 10-year SOFR swap on 13 October, and the 30-year 65bp more. (SOFR is the overnight rate on repo backed by Treasuries; these swaps are priced on it.) The gaps were widest from April to September 2025: about 53–55bp at 10 years in monthly medians, and up to 88bp at 30 years (May). The Treasury's borrowing advisers put the 10-year at 53bp in July 2025. The gaps narrowed from October 2025 to January 2026 and widened again in February and March 2026, the 10-year back to its October 2025 level and the 30-year past it. They have narrowed since. On the same tape, 10-year inflation swaps sat 14bp above the 10-year Treasury breakeven on 13 October: a gap between two different instruments. These are measured from public trade reports, and say nothing about why they moved.
Sources and corrections: each first pass was attacked from outside, and seven corrections came back, C-122 to C-128: C-122 (rate risk per face dollar), C-123 (the Fed's share), C-124 (the household row), C-125 (years five to ten), C-126 (Japan's breakevens), C-127 (Germany's breakevens), C-128 (who absorbed the supply). C-150 (the expected-rate path in the US surveys); R2b (registered funds); ST1 (swap spreads and inflation swaps). Japan's rate-risk range was reviewed from outside on 11 October. Register: CORRECTIONS.md.
Both halves rest on published data, though not every number on them is a measurement (§08 grades them). The join between them, how much a dollar of buying moves a price, is not recoverable on free data. We know exactly why, four times over.
Sources and corrections: C-147 (the Gabaix and Koijen columns); C-117 and C-145 (the payment-day test is furthest from reopening, not closed at every M); C-086 (the Russell cutoff); C-090 and C-091 (the payment-day test was uninformative); C-092 (the mega-firm passive test). Register: CORRECTIONS.md.
A widely repeated claim says America's retirement system has flipped from contributing to withdrawing, ending an automatic bid for equities. We read the source documents.
Sources and corrections: C-112 (the negative run is not an equity measure). Register: CORRECTIONS.md.
Everything load-bearing, graded under the 22 September ruling (§01). Green is measured; blue is an identity, a bound or a hypothesis; red is dead. Besides the four grades, the ledger marks claims that remain open or unresolved (blue) and claims that were retracted or refuted (red). The authority for each line is the findings document it came from.
| Claim | Grade | What settles it |
|---|---|---|
| Earnings share of the price gain since 2015: 71.0% to Dec 2025, 82.3% to Jun 2026 | Measured | Shiller; the old “70–80%” band excluded the second (C-100) |
| …of which the 2017 tax cut is about a tenth of the EPS growth | Measured | 412 firms; 8.9–9.5% by basis (C-098) |
| The earnings yield ended near where it began (−14bp net, after −61bp in 2024) while real yields rose 34bp | Measured | Shiller E/P; FRED DFII10 |
| …and the gap gave back ~17bp in the first half of 2026 | Measured | Earnings caught up |
| 2024’s multiple expansion arrived as drift, not on scheduled news | Measured | FOMC beats earnings head-to-head |
| Megacap depreciation-life extensions lift reported profit | Measured | 5–11% of six firms’ NI growth — real, not decisive |
| ETFs created $3,604bn of shares, counted as corporate equity | Measured | Z.1; matches ICI’s census, on which it appears to be built (C-148) |
| Operating companies retired $541bn of equity net, 2024–26 | Measured | Z.1 non-financial corporates; the −$361bn “net of wrapper” is a blend (C-107) |
| Operating companies retired equity on net: gross retirement $2,979bn (65% buybacks), gross issuance $2,427bn | Measured | Fed EFA 2024:Q1–2026:Q2; net −$552bn (C-108, C-133) |
| ETF buying and mutual-fund selling offset by 82% | Measured | ETFs +$2.46trn vs mutual funds −$2.01trn |
| DC net contributions negative every year since 2013 | Measured | ICI; all assets after transfers — not an equity measure |
| Three of the five large AI spenders generate more operating cash than their capex; signed leases not yet commenced total $1.12trn | Measured | Filings: 3 of the 5 large spenders' operating cash exceeds capex (C-134); leases $1.12trn |
| The US 10-year rose 113bp, almost all of it in real yields | Measured | TIPS +96bp, breakeven +17bp (R1) |
| 30-year TIPS rose more than 10-year | Measured | Treasury daily real yield curve, cross-checked to H.15: +119bp vs +96bp, end-2023 to 18 Sep 2026 |
| The US public absorbed ~36% more rate risk at constant yields | Measured | 10-year equivalents on the end-2023 curve; per face dollar the debt did not get longer (C-122) |
| German 2024–25: the 2033 linker's breakeven fell 21bp as the nominal yield rose 66bp | Measured | Bundesbank daily price-and-yield files, one fixed issue, nominal interpolated to its maturity; attribution bounded by the liquidity premium (C-127 update) |
| German linker breakevens rose by about half or more of 2026's Bund rise | Measured | Finanzagentur daily series: 6.6- and 19.6-year linkers +34–49bp against the 10-year Bund's +68bp; linker breakevens carry liquidity premia (C-127) |
| Banks and public pensions were the largest buyers while the BoJ's holdings fell ¥63trn (overseas investors too from 2025; their purchases exceeded the fall), and banks and the rest of the world while the Eurosystem's fell €751bn (C-128) | Measured | Flow of Funds and ECB holdings by sector; both partitions checked every quarter |
| ~93% of the 2024–26 rise in equity held ($46.9trn) is not transactions; the stock is the $123.7trn held, not listed market value | Identity | Level minus flow = revaluation + other volume changes; uncheckable here (C-109, C-119) |
| Real-rate leg −2.14 P/E pts; premium leg +3.01 | Identity | Sums by construction; 71% of the premium leg is the rate rise (C-102) |
| Buybacks are at most ~7–13% of the S&P 500 price gain since 2015 (a tenth to a sixth of the earnings part) | Bounded — a ceiling | Survivorship pushes every weighting up (C-101) |
| The Fed's share of US duration supply: 4%, at most ~13% | Bounded — a ceiling | Depends on how Treasury refinanced the runoff, which it does not publish (C-123; checked 8 Oct) |
| Japan's public took 11–25% more JGB rate risk, depending on the measure | Bounded | Fixed-coupon JGBs to 31 Aug 2026. 10-year equivalents on the end-2023 curve: +16.0% from Oct 2024, +24.8% from end-2023 (interpolated start). Face × maturity: +10.9% from Oct 2024, +18.2% from end-2023 (outside review, 11 Oct) |
| JGB breakevens did a third to two-fifths of the 10-year's rise | Bounded | 82–97bp of 243bp, by inflation-indexed issue (C-126) |
| Profit + composition at least ~69–75% | Bounded — its complement | The same uncertainty, not a second finding (C-101) |
| The ETF/mutual-fund offset is the same money rotating | Hypothesis | Needs account-level data; none is free (C-111) |
| Long money holds most corporate bonds (measured at sector level); whether it holds most of the AI debt is not measured; money funds hold about $3.4bn under AI names, mostly as repo collateral (a name search; C-138) | Hypothesis | Z.1 sector level; N-PORT census of seven names; money-fund census (C-139) |
| The AI build-out is mostly self-funded, and the financing that matters is off balance sheet | Hypothesis | Inferred from the firm-level counts and the lease stack (C-134) |
| Bank lending funds the AI build-out at scale | Hypothesis | $450bn committed / $150bn outstanding is industry exposure; ~$250bn of the commitments predates the boom (C-115, C-120) |
| Term premium versus expected rates, in any of the three bond markets | Hypothesis | US models and surveys agree on years 1–4 and split on 5–10 (C-125); no free series for JGBs or Bunds |
| Cheaper risk versus better growth | Hypothesis | 36–60% risk share across two rate bases (C-104); needs a third series |
| Who bought — issuer side attributed, holder side unsplit | Open, no free route | Private Fund Statistics scout closed negative |
| Two Fed series for corporate net issuance disagree | Unresolved | +$45.4bn vs +$51.3bn, 2026:Q1 (current vintages; the EFA figure was +$25.2bn on 13 Sep) |
| Household and rest-of-world revaluations fail to reconcile | Retracted | The gap is other volume changes, a term the check left out (C-119) |
| The premium, not rates, held the multiple up | Retracted | The premium is defined against the rate (C-102) |
| Buybacks were fully offset by issuance, 2015–26 | Retracted | 82% of that issuance was private (C-108) |
| Leases are 13× the guarantee stack | Retracted | 4.9× like-for-like (C-116) |
| The payment-day test is closed for good, at any budget | Retracted | Not at every M: on 11.7 years it reopens only above M≈11 (C-117, C-145) |
| A two-fifths / three-fifths risk–growth split | Retracted | Survey premium fell on its bond leg |
| The listed equity base grew | Retracted | Operating companies retired equity (C-081) |
| New deposits capped how much buying money creation could fund | Retracted | Deposit stock is 6.3× the purchases (C-089) |
| 2024 wholesale-repo growth was new cash | Retracted | Reverse-repo handoff (C-085) |
| A fair-value level from holding the premium fixed | Retracted | Moves 10.9%–43.1% with the baseline (C-078); its CAPE twin too (C-103) |
| A borrowed multiplier converts flows into price impact | Retracted | No central value; checked estimates span about 1.5 to 11 (C-077, C-145, C-147) |
| The Russell cutoff recovered an elasticity | Retracted | First stage never assigned treatment (C-086) |
| The payment-day test rejected the mechanism | Retracted | Uninformative — ~386 years needed at the multiplier the test was sized on, M≈1.9 (C-090/091) |
| The mega-firm passive test rejected the mechanism | Retracted | Uninformative — saw 0.48× of its target (C-092) |
| US duration supply outran face value | Retracted | Par-weighted it did not; the gap was higher coupons (C-122) |
| Issuance alone supplied the extra duration; the Fed's share was 4% | Retracted | The Fed's share is up to ~13% on a refinancing counterfactual (C-123) |
| A survey split showed the Bund's rise almost entirely real | Retracted | Market breakevens moved in 2026 (C-127) |
| The Dutch pension transition sells long-duration bonds | Refuted | E1 scout: a relabelling; the DC book keeps the allocation |
Several gaps have closed since the first version: earnings quality, which part of supply moved, whether deposits funded the buying, the Dutch transition, the join itself. What remains:
Work on currencies and commodities began on 12 October with a first measurement map: who holds Treasuries, how oil and gold flows moved, and what funds did in futures. Four tests were fixed in advance; none shows cause.
| Test, fixed in advance | Result | What it shows |
|---|---|---|
| Do fund positions in oil lead the price? 731 weekly reports for WTI, the US crude benchmark (September 2009 to March 2024, futures prices), and 819 for Brent, the North Sea benchmark (2011 to September 2026, spot prices as a stand-in) | A change in fund positions was not followed by a detectable change in the next week's price. No lead larger than a correlation of 0.083 (WTI) or 0.105 (Brent) was detected. | No lead detected; descriptive only |
| Gold and the US 10-year real yield, monthly, 2006–2026 | In months when the real yield rose one point more, gold's monthly average price was about 7.7% lower, holding the dollar fixed (95% band 4.6% to 10.6%). The link looks weaker since December 2021, but those windows are too short to tell from no change. | Association only |
| The yen and the US–Japan two-year yield gap, over the 118 weeks since March 2024 that the test keeps (intervention weeks left out) | Week to week, a narrower gap went with a stronger yen. Over those weeks the yen fell about 13% against the dollar in log terms (a continuously compounded measure, slightly different from an ordinary percentage), about 4% net if the intervention weeks are included. The gap does not account for that weakening, and the test cannot say what did. | Association only |
| How closely the US, German, UK and Japanese 10-year yields move together, weekly, 2024–26 against 2018–19 | No reliable change. One common factor accounted for 0.71 of their weekly variation in 2024–26 against 0.78 in 2018–19, a drop too small for the test to tell from zero. | Association only |
Who holds Treasuries. Over the year to 2026:Q2, against the year before, the rest of the world's share of net Treasury issuance fell from 44% to 12%. Money funds took 28%, all of it in late 2025 (they were net sellers in 2026), and the Fed turned buyer of bills in 2026 (Z.1; offshore hedge funds sit in its household row). Private foreigners are still buying US securities while official holders shrink their Treasuries.
Oil and gold flows. Gold rose alongside official buying, bar-and-coin demand and ETF inflows, while managed-money futures length ended where it began. (Managed money is the futures reports' category for hedge funds and other money managers; length is their net long position.) Oil's rise ran with large stock draws. Unreported central-bank gold is not measurable on free data.
Fund positions in oil. In NYMEX WTI and ICE Brent, funds' futures length rose over 2026 in swings rather than one climb. The largest builds came in January and February, before the price moved, and from July to mid-September as prices rose. Brent also added through the March surge and was then cut from its 31 March peak to early July. WTI's position swung between about 64,000 and 100,000 contracts from March to early July. Both were cut after the 15 September high. Since January 2024, fund positions moved with the price week by week, more closely in Brent than in WTI. Over the longer test samples the two are similar (same-week correlation about 0.45 for WTI and 0.42 for Brent). They did not lead it (first test above). In the four weeks to 6 October managed money halved its WTI futures net long, from 111,731 to 53,598 contracts of 1,000 barrels; counting options too, it fell by a quarter, from 139,339 to 104,491. Over the same four weeks it cut its Brent futures net long by 28%, and by about a third from the 15 September high (ICE). The flows are measurable.
Hedging cost. Part of the cost of hedging is now measurable too. On the public swap tape (DTCC) the 1–5-year premium for borrowing dollars against yen roughly halved in two years, to about 0.2 percentage points; against the euro and sterling it is near zero. Add the gap between the two countries' short-term interest rates and you get the market cost of a hedge renewed every three months, at each month's prices. For a yen-based investor that cost was about 3.9 percentage points a year in 2025 and 3.0 in 2026, and about 94% of it was the interest-rate gap, the rest the dollar premium. After that cost, a hedged 10-year Treasury carried 1.0 to 1.4 points less than a 10-year Japanese government bond in each of the 18 months with enough swap trades to measure, March 2025 to September 2026 (an average of 1.27 points in 2025 and 1.12 in 2026 to September). For a euro-based investor the hedged Treasury carried more than the euro-area 10-year yield from October 2024 to January 2025, and less from March 2025 (by up to 0.79 points, in July 2025). Since June 2026 the two have been about the same. (Three months in 2025 have too few trades to measure.) Our figures sit a median of 7 to 12 basis points (0.07 to 0.12 percentage points) from published hedged yields (BlackRock, TD Economics, Ginnie Mae), though one of the ten points compared was off by more than 25. This measures an incentive. It does not show what led any investor to buy, sell or hedge. How much foreign demand is hedged is still not measured.
Sources and corrections: C-095 (stock-based compensation and R&D are moot); C-102 (the compression is the rate rise restated); C-104 (36–60%). Currencies and commodities: E-021, the 12 October measurement map. Its rates and FX claims were reviewed from outside on 13 October and one fell (C-136); its commodities claims were reviewed the same day: one fell outright (C-141), two more were withdrawn after being weakened (C-142), and six were reworded. C-124 (offshore hedge funds in the household row); C-146 (swings in fund length); the four tests: E-021 builds, 18 Oct; hedging cost: ST1, and the hedged Treasury yield, 19 Oct (data/hedged_ust_yield). Register: CORRECTIONS.md.
The third-quarter registered-fund file shows Beignet flat at about $9.7bn, 36% of the $27bn deal. The money-fund filings for 30 September show the conduit channel has not stopped (named conduit families $51.2bn, from $49.7bn in July).
Gates: rebuilding the channel map dealer-first stays on hold while the conduit channel runsThe principal reads the verbatim goal statement, not a summary of it, and says whether it still describes the intent. It is the only check that has ever caught goal corruption here. The principal's answer is awaited.
Owner: the principalWritten down before the 16 October release, after Japan's finance ministry intervened with ¥15.4trn between 30 July and 26 August. The Treasuries US custodians record for residents of Japan fell from $1,103.9bn to $1,068.3bn: $35.6bn, or $40.4bn after taking out the $4.8bn a 10-year Treasury would have earned over August. That is inside the range set in advance (−$190bn to +$60bn), which was wide because earlier interventions showed no stable ratio. The narrower guess, a fall of $50–100bn, missed: holdings fell by less.
Showed: recorded holdings fell about 0.4 dollars per dollar of intervention, inside earlier episodes' spread; an accounting comparison, not proof that sales paid for itA current-report filing (an 8-K) for a draw on the $17.5bn term loan by 30 September was due by 6 October. None appeared in the filings listed to 6 October, at the 8 October check. No filing does not prove no draw, so the question stays open until the quarterly filing, due late October or November.
Tests: whether the lease stack needs bank money at the marginThe six filers' leases signed but not yet commenced, re-read by hand because the $1,122.9bn has no machine-readable tag; the five hyperscalers' free cash flow; and Alphabet's energy-equipment guarantees.
Tests: the self-funded verdict and the off-balance-sheet totalThe Treasury's quarterly announcement of how large its coming bond auctions will be, at which maturities, and how much long-dated debt it will buy back. It is scored by a rule set in advance. The 10- and 30-year yields must both move more than 8bp that day, in the direction the supply news points: up for larger auctions or smaller buybacks, down for smaller auctions or larger buybacks. On Wednesdays without a Fed decision or an inflation release, each yield stays within 8bp either way at least nine times in ten. Only two statements since November 2023 raised auction sizes, in November 2023 and January 2024; both came on Fed decision days, and yields fell on both. No statement has passed the test on a clean day, and on 5 February 2025 both yields fell more than 8bp with auction sizes unchanged, so a large move alone proves nothing. If neither auction sizes nor the long-end buyback limit changes, there is nothing to score. This statement comes the day after the US election, so even a hit stays a hypothesis.
Tests: whether news about Treasury supply moves yields on the dayWritten down before the six-monthly derivatives release of the Bank for International Settlements (BIS). The BIS publishes dollar FX forwards and FX swaps as one combined series. The prediction: its notional amount outstanding will be 11% to 18% higher at end-June 2026 than at end-June 2025 ($75.7trn), that is $84.0–89.3trn. The last two June-on-June changes were +8.2% (2024) and +14.5% (2025).
Tests: whether the programme's proxy for off-balance-sheet dollar borrowing keeps growing (it includes outright forwards, which are not borrowing)The third-quarter Financial Accounts (about 11 December), which refresh the bond-buyer table and the wholesale share; Oracle's quarterly filing, which shows whether its matured lessor guarantee was repaid or rolled over.
The turn-of-year prints for SOFR, the Fed's reverse repo, sponsored repo and commercial paper, compared with the June and September turns. Prints only: the two sentiment runs found almost no chatter, so a third would add little.
These are credences, not proofs: surviving an attack is a fact about the attack. High on the earnings share and on the wrapper reclassification. Both are measured (the earnings share's decomposition is an identity, and its split is basis-sensitive by about 7%), and the ETF figure matches ICI’s own census. Medium, and lower than it reads, on the premium story. What is measured is that the earnings yield did not follow real yields up; three premium constructions agree on direction for a shared structural reason, not independently. Low, and explicitly bounded, on causation: there is no aggregate estimate, and §06 says why four times over. On rates, the levels and holder tables are measured, the breakeven splits hold only within their stated ranges, and no term-premium reading is measured at all.
Would. Earnings did most of the work, and it is mostly real profit. The multiple held up while real yields rose, and has partly given that back. Operating companies retired equity on net while the ETF layer grew. ETF buying and mutual-fund selling largely offset. No measure shows collateral re-use rising before the price rise, and new shadow money in 2024 was about a tenth of households' equity purchases.
Would not. Any multiplier. Any statement that flows caused a specific share of the repricing. Any split of the household line by wealth group or buyer type. A point estimate for buybacks' share. Anything about retirement flows after 2023. That the premium was “cheaper risk”. That bank lending funds the build-out at the scale of its headline figure.
Almost none of the corrections registered up to C-119 were arithmetic errors: nearly every figure in them reproduced exactly. They fall into five shapes, and each now has a guard.
The largest corrections — the ETF wrapper, the risk–growth split — came from adversarial review by a second model, not from internal reading. Internal review reads the same reasoning that produced the error and agrees with it. Outside attack stays the highest-yield activity in the programme.
Several corrections were registered correctly while a second copy of the dead claim lived on elsewhere — in the synthesis, in a pointer file, and on this page. An edit made in place without a correction number is invisible to the checks. Every copy of a number is a place it can go stale: figures now live in the findings note that owns them, the answer note carries no headline of its own, and this page is scanned before it is republished.
Three designs returned nulls that were written up as rejections, when none could have detected the effect it was testing for. The minimum detectable effect is now computed before a design is commissioned — from a realised clustered standard error, not a formula that counts rows — and a kill switch on an underpowered test no longer counts as a result.
A decomposition’s “sum check: exact” was exact because one term had been solved for as the remainder; measured directly, that term was 7.6% different. A validation reporting “every year within tolerance, worst year 2019” was an algebraic identity. The corrections gate itself went blind to every correction numbered above 99 and kept printing green. A check that cannot fail is not a weak check — it is the absence of one. Residuals are now measured a second way where possible, and the gap is reported.
Four times in two days a careful source document hedged a finding and the one-line summary standing on it did not — the shadow-money verdict, the premium, the pension flows, the bank-credit figure. The research was sound; the compression was not. The principal’s ruling of 22 September — everything is a hypothesis until proven — now grades every claim, and this page carries the grades.
The guard rails are these. The principal's statements are recorded verbatim and never paraphrased. A corrections register carries machine-enforced bans on dead phrasings, and is read by its pass lines rather than its exit code. Every outside number is re-derived locally before it is believed. Two agents hand over after every piece of work with an integrity stamp, so that neither works from a stale copy of the record.