Cornering the Curve: What SOMA Securities Lending Reveals About Treasury Collateral Scarcity

In May 1991, Salomon Brothers gained control of at least 86% of the on-the-run two-year Treasury note. They bid from their own account and from clients' accounts at the same auction, working around a rule that capped any single dealer at 35% of an issue. Short sellers who needed that note to settle their trades had nowhere else to go. Repo rates on the issue fell sharply into negative territory, and Solomon benefited greatly as they were getting paid to take the other side of the repo transaction. (Jordan & Jordan, 1996, Journal of Banking & Finance).

That episode is the clearest illustration of what happens when a single entity controls too much of a specific Treasury issue.

Welcome back to the macrodispatch. This entry covers a new dataset I've spent the past year building: SOMA securities-lending data cross-referenced against Treasury auction records. If you find value in it, donations help keep it free; if you use it professionally, even more so. I first ran into this data while talking with R. Jay Kahn a couple of years ago, and this year I finally went down into the weeds.

Why does the Fed lend bonds it doesn't need to lend

SOMA is the Federal Reserve System's Open Market Account, where the Fed parks the dollar-denominated assets it acquires through open market operations. It holds a lot of paper. The Fed isn't really a trader; it shows up at auctions mainly because old securities mature and the cash has to go somewhere, then it holds the new ones until they mature too.

But holding to maturity creates a problem for everyone else. Bonds locked in SOMA aren't bonds available for delivery, short sales, or repo. So the Fed lends them back out through its securities lending facility, giving primary dealers fresh collateral for delivery-versus-payment repo, including the kind of repo that lets a dealer short a specific issue.

This facility is the best public proxy available for specific repo rates, updated daily using SOMA lending auction data. When a particular CUSIP isn't available in the secondary market and demand is high, that issue goes "special": its repo rate drops below the general floor, sometimes sharply. That's the mechanism Salomon exploited in 1991, and it's the same mechanism this dataset measures, just under far less extreme conditions today. The Fed's balance sheet has grown enormously since the 1990s, through the shift to an excess-reserve system and the 2008 crisis, which mechanically dilutes any single dealer's ability to repeat what Salomon did.

That same balance sheet growth created a different dynamic, though. As securities mature, the Fed has to redeploy the cash, and it does this through the SOMA reinvestment mandate, placing non-competitive bids at Treasury auctions sized to whatever is settling that day. Because maturities are spread across the calendar, the Fed shows up at nearly every auction. That regularity is what makes cross-referencing SOMA holdings against auction data worthwhile: you can watch, issue by issue, how much of each new bond the Fed locks away before the private market ever gets a crack at it.

The shape of the dataset

This covers 668 unique on-the-run CUSIPs across the 2Y, 5Y, 10Y, 20Y, and 30Y tenors, from 2010 to today. Across all securities in the lending facility, the Fed has accepted $97.8 trillion in cumulative par value over that period. That's not a stock figure but a flow one: the daily churn of collateral going out and coming back.

The lending facility's scope is even larger: over 2,100 distinct CUSIPs have appeared in lending operations, including bills and other short-dated paper. This piece focuses on the coupon curve, 2Y through 30Y, because that's where on-the-run dynamics matter most, collateral scarcity bites hardest, and the relationship between SOMA holdings and dealer positioning is clearest.

How much does the Fed actually hold? As of the latest data: roughly $10.2 billion of the current 2Y on-the-run, $10.3 billion of the 5Y, $10.0 billion of the 10Y, $5.9 billion of the 30Y, and $2.4 billion of the 20Y (reintroduced in 2020 and still building up its history). The QE era is unmistakable in the 10Y line: holdings of a single CUSIP ballooned past $36 billion during 2020-2022, a direct consequence of the Fed's pandemic-era purchases. As QT took hold and those issues rolled off, the numbers came back down.

Raw dollars only tell half the story, though. What actually matters for market function is how much of an issue's float the Fed controls: SOMA holdings divided by the original auction size.

Here, the picture makes an interesting point. At the peak of QE, the Fed owned 61% of the on-the-run 10Y note (CUSIP 91282CDY4, May 2022), over $36 billion of a $59 billion issue. The 20Y wasn't far behind, at nearly 45% of the May 2021 on-the-run (912810SW9). For every $100 of face value outstanding on those issues, the Fed had locked away $45 to $61, leaving the rest for the entire private market to trade, short, and repo against.

Today, post-QT, those shares have normalized. The 10Y and 30Y both sit around 19%, the shorter tenors closer to 13%. But the historical swing is the point: when the Fed's share moves that much, the amount of collateral actually available to dealers moves with it, mechanically and immediately.

Summing SOMA holdings across all active on-the-run tenors on a given day, divided by total first-auction par, gives a single number: what share of the on-the-run market does the Fed hold right now. It's a direct proxy for the cumulative weight of non-competitive reinvestment bids. High readings mean the Fed is absorbing a larger slice of each new issue, mechanically leaving less for competitive bidders and the secondary market. Low readings, like the current QT-era numbers, mean the private market has reclaimed that float.

What do dealers actually pay to borrow? The most market-relevant number in this dataset is the lending rate, the fee the Fed charges when it lends out a CUSIP. It's a direct read on scarcity: the higher the rate, the more desperate dealers are for that specific bond since the baseline is 5 bps, it's expected that the median is 5 bps since not all the days that a issue will be on special (that would kind of deafeat the point of the name).

The tails tell the sharper story. The single highest on-the-run special rate in the entire dataset belongs to the 20Y: 684 bps, hit on September 27, 2024, a quarter-end. The 10Y peaked at 511 bps on March 13, 2020, the week COVID panic made Treasuries themselves illiquid. The 2Y spiked to 409 bps on September 29, 2022, another quarter-end. Three different tenors, three different years, the same calendar trigger showing up twice. The 20Y's median rate sits at the same 5 bps floor as every other tenor, yet its P95 jumps to 203 bps, almost ten times the 5Y's. Thin float plus infrequent issuance equals a fat tail whenever demand spikes, and liquidity is thinner for the 20 years since it was only reintroduced in 2020 after a 34-year hiatus. The smaller amount of data makes it more prone to outliers, as we can see from the higher average.

Although the 20Y commands the highest average special rate at 31.7 bps, the 10Y follows at 20.8 bps. The front end runs cooler, with the 2Y at 13.3 bps and the 5Y at 12.8 bps, which makes sense given how often new issuance refreshes its float. The 30Y, despite its long duration, averages just 6.8 bps.

Why does the rate spike where it does? The SOMA auction sits at the bottom of a collateral sourcing hierarchy. Dealers exhaust internalization and private repo first, through bilateral DVP, securities lending from real-money holders, and GCF, and only reach the noon SOMA auction once those channels run dry or price above the Fed's rate. By the time a SOMA special rate spikes, the private market for that CUSIP is already deep in special territory. The auction is the visible terminus of the squeeze, not its starting point.

That logic explains the cliffs as much as the spikes. As we see in the example beloe in detail, when the 10Y special rate collapsed from 61.9 bps to the 5 bps floor in mid-June 2026, the Fed hadn't changed anything. Delivery obligations cleared, internalization opportunities normalized, and the stack drained from the top down. The SOMA rate works as a proxy precisely because it sits at the bottom of that stack: when it moves, everything above it has already moved; when it normalizes, the squeeze has already resolved.

Quarter-ends, reliably

Quarter-ends are when balance sheet constraints bind hardest. Dealers shrink their repo books, prime brokers call in collateral, and the scramble for specific Treasuries to settle regulatory reporting obligations intensifies. The SOMA lending data captures this cleanly: across the full 2010-2026 sample, the average special rate within seven days of a quarter-end runs 23.5 bps, against 14.8 bps on normal days, a 59% premium that shows up reliably enough to set a clock by.

The effect concentrates almost entirely in on-the-run paper. Splitting the data by run status (OTR, first/second off-the-run, and everything else) makes the contrast stark:

OTR rates jump 47% at quarter-end, from 10.6 to 15.6 bps average, while OFR barely moves, from 6.0 to 6.6 bps. The P95 gap is even louder: OTR's tail nearly doubles, from 17.7 to 34.2 bps, while OFR's creeps from 7.2 to 9.9 bps. Dealers scrambling for collateral at reporting dates are chasing the benchmark issue specifically, not whatever rolled off the run most recently.

The volume numbers back this up at the aggregate level, too. OTR issues make up just 3.8% of all lending observations in the dataset but carry an average rate double the catch-all "Other" bucket, and a P95 nearly four times higher. Of the $97.8 trillion in cumulative lending volume since 2010, on-the-run issues account for $12.4 trillion, about 13% of the flow, yet generate a wildly disproportionate share of the actual stress as we can observe the large amount of outliers in the box plot below.

June 2026: a squeeze in real time

The historical patterns above span fifteen years. The dataset updates daily, though, which means you can watch the pattern happen instead of just reading about it after the fact. Let's set an example with data fresh off the press:

The on-the-run 10Y, CUSIP 91282CQQ7, ran hot through the first half of June. Its special rate hit 61.9 bps on June 3rd and stayed above 20 bps for the next nine trading days, with dealers borrowing over $6 billion of it per day. Then, between June 13th and June 15th, it fell off a cliff, back to the 5 bps floor, with accepted par collapsing from $6.4 billion to $52 million by June 23rd. Intense demand followed by sudden normalization is the signature of a settlement-driven squeeze resolving once the delivery obligation clears, exactly the mechanism described above.

The 20Y (912810UV8) showed a milder version of the same shape, touching 14.4 bps on June 1st before fading back to the floor over the following weeks. The 2Y, 5Y, and 30Y stayed flat at 5 bps essentially the whole month.

No other coupon CUSIP broke above the 5 bps floor all month, on-the-run or off. The 10Y alone accounted for $131 billion in accepted par, nearly as much as the other four tenors combined, at an average rate that was six times the next-highest CUSIP. June's collateral stress wasn't a curve-wide event. It was one bond.

Zooming out across 2025 and 2026, that pattern of clean quarter-end spikes repeats on schedule: March and June 2025 both show OTR rates jumping above the OFR and Other baselines, September 2025 pushed OTR to nearly 17 bps in the week of the 8th, and December's year-end was the sharpest of the year, with OTR averaging 34.3 bps that week, driven by a single 456 bps day. Outside those windows, 2026 has been calm, with OTR hovering near the 5–6 bps floor. OFR and Other barely move through any of it, which is itself the point: collateral stress in this market is almost entirely an on-the-run phenomenon.

What does this data tell us?

This is another piece of the puzzle. It is a saying around the market that the repo market is the largest and least transparent market in the world. But if you actually dig into the data, you can pick up a pretty decent color of what's happening: is everything working fine, or is there some size, some real sense of stress building?

With this data set, instead of relying on secondhand accounts and the whispers they hear, we can see, across the curve, what's happening with T+1 timing for any CUSIP. It's a genuinely interesting way to see what's going on — and a free way to monitor what's actually happening, instead of relying on a Twitter feed telling you some liquidity specialist heard from a friend that things might be bubbling up in the repo market. We don't work with that. We work with facts: what's the historical pattern? Can we visualize it? Real reporting on whether it's on-the-run or off-the-run.

If you like this kind of content, share it with a friend.

Thanks to RJ for the tip on this dataset, high granularity, and badly underused public data.

Data sources: NY Fed SOMA securities-lending operations; Treasury auction results. All figures are par value. Dataset maintained and cross-referenced by macrodispatch.