SOFR's 99th Percentile Reached the Target Range Ceiling on Two Days, Both Month-Ends

On two sessions between July 1 and September 2, 2026, the most expensive slice of the secured overnight funding market traded at or above the ceiling of the Federal Reserve's policy corridor. On July 31 the 99th volume-weighted percentile of the Secured Overnight Financing Rate printed at 3.75 percent. On August 31 it printed at 3.77 percent. Those were the only two sessions in a 45-business-day window where that figure reached 3.75, and both of them fell on a month-end.

The headline rate did nothing comparable. SOFR itself moved between 3.53 and 3.68 across the same window and stayed inside the target range on all 45 days. A reader watching only the published median would record two quiet months. The distribution underneath that median was not quiet, and the dates on which it stopped being quiet were calendar dates rather than economic events.

The question this piece takes up is narrow. When the middle of a distribution sits still and its upper tail does not, which of the two is carrying information, and what would have to show up in September to separate a calendar effect from a level shift? Every difference, count and average below is arithmetic performed here on published daily values. Neither the Board of Governors nor the New York Fed publishes any of them as a series.

The Corridor the Rate Has to Sit In

Four numbers bound the box that overnight rates are steered into, and the Implementation Note issued with the July 29, 2026 FOMC decision sets each of them. It directs the Desk to “Undertake open market operations as necessary to maintain the federal funds rate in a target range of 3-1/2 to 3-3/4 percent.” It directs the Desk to “Conduct standing overnight repurchase agreement operations at a rate of 3.75 percent.” It directs the Desk to “Conduct standing overnight reverse repurchase agreement operations at an offering rate of 3.5 percent and with a per-counterparty limit of $160 billion per day.” And the Board of Governors voted to “maintain the interest rate paid on reserve balances at 3.65 percent, effective July 30, 2026.”

Stacked, those give a floor at 3.50, an administered rate for bank reserves at 3.65, and a ceiling at 3.75 that is simultaneously the top of the target range and the rate on the Fed's standing overnight repurchase agreement operations. The 25-basis-point corridor is the frame. Everything below is about where inside it, or outside it, actual transactions printed.

On August 31 the effective federal funds rate printed 3.63 and SOFR printed 3.68. The unsecured benchmark sat 2 basis points below the rate banks earn on reserves; the secured benchmark sat 3 basis points above it. EFFR held at exactly 3.63 on every session from August 25 through September 1 while SOFR ranged from 3.64 to 3.68 over the same six days. Size is part of the reason the two behave differently. On August 31 SOFR rested on $3,056 billion of transactions and EFFR on $105 billion, a ratio of about 29 to 1 on that date, computed here from the two published volumes.

The Average Is the Least Informative Number in the Series

Counting sessions is more useful than averaging them. SOFR closed at or above the 3.65 IORB rate on 10 of August's 21 business days, against 5 of July's 22. That is a move from roughly a quarter of sessions to roughly half. Both counts come from the daily series reproduced in the source file for this piece; neither is a published statistic.

The width of the band drawn above is the second derived measure. The gap between the 99th and the 1st volume-weighted percentile averaged 12.90 basis points across August's 20 non-month-end sessions and reached 17 basis points on August 31. That is 4.1 basis points wider than the ordinary session, about 32 percent. On July 31, the other month-end in the window, the 99th percentile reached 3.75 exactly.

Volume moved with it. August 28 carried $2,808 billion, the month's lowest reading. Three calendar days later, on August 31, the figure was $3,056 billion, an increase of $248 billion or 8.8 percent in a single session. More cash was changing hands at a wider spread on the day the calendar turned.

One structural point belongs here, because it decides how much weight the band can bear. The distribution is trimmed at the bottom and not at the top. The New York Fed's methodology page states that “In order to mitigate the influence of 'specials' transactions on the measurement of the general cost of financing, 20 percent of the lowest-rate transaction volume from the DVP segment is removed each day.” The removal is described in terms of transaction volume rather than a count of trades, and the New York Fed describes the published percentiles as volume-weighted in the same way. A band that widens can therefore widen for reasons that live entirely in the upper tail, and a symmetric reading of it would be wrong.

The same series argues against reading any single print as a regime. On July 9 SOFR printed 3.53, which is 12 basis points below IORB and 3 basis points above the ON RRP offering rate. Nothing in the corridor changed that week. The full window therefore spans 15 basis points, from 3.53 to 3.68, with the four administered rates fixed the entire time. This piece does not offer a cause for the July trough, because none was confirmed in a primary source.

What a Reporting Date Does to a Secured Rate

The mechanism connecting a date to a rate is not inferred here. It is stated by the manager of the System Open Market Account. In remarks before the Money Marketeers of New York University on March 26, 2026, Roberto Perli listed what an ample-reserves regime should look like, including that repo reference rates “should not deviate too much from IORB, on average, but show a moderate amount of volatility, especially around ‘high-pressure’ days in the repo market.” He identified those days in a footnote: “These include days with high payment flows due to Treasury securities settlements or tax payments, and balance sheet reporting dates, when temporarily tighter balance sheet constraints reduce dealer repo intermediation capacities.” A second footnote adds that “Usage of both ON RRP and SRP operations may increase around reporting dates due to temporary reductions in dealer repo intermediation capacity and other frictions.”

The chain has four links and each one is observable. Cash lenders arrive with overnight balances. They choose between the Fed's reverse repo facility at 3.50 and dealer repo. Dealers take the other side using balance sheet capacity that is not constant across the month. What SOFR reports the next morning is the volume-weighted result of those choices. When the third link narrows on a reporting date, the price of the fourth moves and its dispersion widens, which is what the August 31 figures show.

The same speech places the reserve-supply question in a specific history. Perli noted that “When it [the FOMC] judged that reserves declined to the ample range in December, it instructed the Desk to begin reserve management purchases (RMPs) to maintain reserves within that range.” He also observed that “Repo reference rates rose relative to IORB and sustainably printed above EFFR.” Both statements describe 2025 - the judgment named there was made at the December 2025 meeting - and are cited here as background, not as a description of March 2026 and not as a description of September.

One limit on the reading is worth stating plainly. SOFR is not a single venue. The methodology page notes that “The SOFR includes all trades used in the BGCR plus data on transactions cleared through the Fixed Income Clearing Corporation's Delivery-versus-Payment (DVP) repo service.” Which of those segments produced the 3.77 print is not identified here, because the published data do not break the percentile down by segment.

Three Branches for September

These are branches with observable discriminators, not forecasts. Each one is distinguished by a specific number arriving on a specific date.

Branch one: the date, not the condition. September 30 is a quarter-end, which is a heavier reporting date than an ordinary month-end. If the pattern is calendar-driven, the September 30 print should show a wider band than the 12.90 basis point ordinary session and a 99th percentile at or through 3.75, and the first business day of October should return to the 3.62-to-3.66 range that described August's other sessions. A tail that spikes and then immediately unwinds is a balance sheet effect.

Branch two: the level is drifting. The share of sessions closing at or above IORB went from 5 of 22 in July to 10 of 21 in August. If September lands above half and the mid-month sessions rather than the turn are doing the work, the calendar reading weakens. The distinguishing evidence is where in the month the elevated prints fall, not how high the peak gets.

Branch three: the corridor is adjusted. The FOMC meets in mid-September. A technical change to the 3.65 IORB rate, or a change in the pace of reserve management purchases, would move the whole distribution rather than its upper tail. That is the cleanest case to identify: the level shifts, the band width does not, and the count of sessions above IORB resets against a new reference rate rather than the old one.

What Would Invalidate This

Two observations are not a base rate. July 31 and August 31 are two data points, drawn from a window of 45 business days that contains exactly two month-ends. A pattern that holds on both available instances of a thing that has occurred twice is not evidence of much, and the honest description is that the sample is too small to distinguish a rule from a coincidence.

The window also contains its own counterexample. On July 9 SOFR printed 12 basis points below IORB with no policy change and no reporting date, which shows the series can travel a long way on non-calendar grounds. If mid-month sessions can move that far, a 3-basis-point month-end move is not automatically structural.

The dispersion measure has a known asymmetry, described above: the low tail is trimmed by rule and the high tail is not. Anyone comparing band widths across days is comparing a trimmed lower bound with an untrimmed upper one, and a change in the composition of the highest-rate volume will move the number without any change in the general cost of financing.

Finally, the unsecured market saw none of this. EFFR sat at 3.63 through the entire late- August stretch. A reader whose exposure prices off the unsecured benchmark would find nothing in these two months worth acting on, and would be right about their own position.

What to Watch Next Week

• Daily SOFR prints and the 99th percentile. Whether the percentile returns to the 3.70-to-3.74 area that described ordinary August sessions, or holds nearer 3.75.
• The share of sessions at or above 3.65. August ran 10 of 21. A first full week running above half would argue that the level, not the calendar, is moving.
• The mid-month settlement dates. Perli's list names Treasury securities settlements and tax payments alongside reporting dates as separate categories. A widening that arrives away from the turn has a candidate cause on that list rather than in the calendar.
• The FOMC statement and Implementation Note. Whether IORB stays at 3.65, whether the standing repo operations rate stays at 3.75, and whether the reserve management purchase language changes.
• The SOFR-EFFR gap. It ran 1 to 5 basis points from August 25 to September 1, widest on August 31. A gap that stays at the wide end away from a turn is the clearer signal.

Concrete Framework

Six quantities, all computable each morning from the New York Fed reference rates and the current Implementation Note. None of them is published as a series, so each has to be constructed.

1. SOFR minus IORB, in basis points. Positive readings are the ones to count. August produced 10 of them in 21 sessions.
2. The 99th percentile against 3.75. Record the date whenever it reaches the ceiling. Two such dates exist in the July-to-September window, and both are month-ends.
3. The 99th minus the 1st percentile. Compare against the 12.90 basis point ordinary session average rather than against the previous day.
4. Transaction volume. A wider band on higher volume is a different event from a wider band on thin volume. August 31 carried $3,056 billion against $2,808 billion three days earlier.
5. SOFR minus EFFR. This separates a secured-market event from a money-market-wide one. The unsecured rate held at 3.63 while the secured rate moved.
6. Position in the month. Tag each elevated print as turn or non-turn before drawing any conclusion from it. The distinction between branch one and branch two rests entirely on this tag.

The corridor has not moved since July 30. What has moved is where inside it the money actually traded, and on two dates, where outside it a slice of the money traded. That is a smaller claim than a funding-stress narrative and a larger one than nothing.