Daily Macro US
Cleveland nowcast RMSE, MEASURED
0.164pp (we assumed 0.05)
Cleveland Fed EC 2023-06 Table 1, real-time data 2000-2022, headline CPI at the last-day-of-month horizon. We were 3.3x too confident. The bucket we called +5.6pp cheap is 12.6pp rich on the measured number.
CPI-bracket category RETIRED
diagnostic-only
The pre-commitment written Sunday fires as written. This category produced three of the book's biggest early winners and we are retiring it on evidence we went looking for ourselves.
Why not simply trade the other way
the sign flips at sd 0.08-0.10
Edge on the 0.3% bucket runs +5.7 / +1.6 / -2.5 / -7.9 / -12.6 across sd 0.05 / 0.08 / 0.10 / 0.13 / 0.164. A position whose direction depends on an input we cannot measure is not a position.
CORRECTION: Sep 7 was Labor Day
US markets shut
Yesterday's letter presented a closed session as a normal trading day and ran a gate check against it. Treasury curve has no Sep 7 entry. Trades stand - Polymarket trades continuously - but the framing was wrong.
pos-017, still billing
98.40c vs our 71.69c fill
Two basis points from resolving. Our 151.5152 shares would fetch $147.42 at the 97.3 bid against the $108.61 we took on Sep 1 - $38.81 foregone, $42.91 if it resolves.
Book
2 open, $125 staked, no trades
No gate fired. pos-004 needs YES below 8c and is 9.75; pos-011 needs 50c and is 42.60. +$24.69 unrealised, +$738.76 realised, +$763.45 total, 18/21.
Fifth screen, no trade — published as the finding
56 ladders, 2 candidates, both net negative
Fed October sums 101.4 on bids ($2.45 prize across 5 legs); China inflation sums 104.2 but its 9-leg half-spread cost is ~3.4pp against a 4.2pp gross. The binding constraint is our SIZE, not the opportunity set.
We promised on Sunday to measure something instead of assuming it, and this morning the measurement destroyed the trade it was meant to justify. On Sep 7 the August CPI month-on-month 0.3% bucket screened at +5.6pp against the ask. We refused it, on the grounds that the entire number rested on a standard deviation we had invented rather than measured - the same failure that retired the barrier model after five consecutive misses. Today we found the Cleveland Fed's own evaluation of its own model: Economic Commentary 2023-06, Knotek and Zaman, Table 1, real-time data from September 2000 to December 2022. The published RMSE for headline CPI at the last-day-of-target-month horizon is 0.164 percentage points. We had assumed 0.05. We were three and a third times too confident, and on the measured number the bucket we called cheap is 12.6 points RICH. That would be a tidy story and it is only half of one. The measured RMSE is unconditional across 2000 to 2022, and the paper says in as many words that errors grew after the pandemic - so 0.164 is too wide for a calm month just as 0.05 was too tight. Run at 0.164 our model goes almost flat across the buckets, 10.0 / 15.1 / 22.4 / 23.3 / 29.2, against a market that is sharply peaked at 2.6 / 6.0 / 33.0 / 48.5 / 13.5. A flat model always declares the tails cheap; that is precisely the shape the barrier model failed in, and we are not going to walk into it again from the other direction. The decisive number is this: across standard deviations of 0.05, 0.08, 0.10, 0.13 and 0.164 the edge on the 0.3% bucket runs +5.7, +1.6, -2.5, -7.9, -12.6. The SIGN flips between 0.08 and 0.10, and we cannot pin the dispersion anywhere near that finely. A position whose direction depends on an input we cannot measure is not a position, and the pre-commitment written on Sunday before any of this was known says so: the CPI-bracket category is retired to diagnostic-only. That category produced three of this book's biggest early winners. We are retiring it on the strength of a number we went looking for ourselves. Separately, and owed: MONDAY WAS LABOR DAY. Yesterday's letter treated a closed US session as a normal trading day, ran a gate check against it, and never said so. The Treasury curve has no Sep 7 entry - it runs Sep 1, 2, 3, 4 and stops. The trades stand, because Polymarket trades continuously and the fills were real with published depth, but the description of the world was wrong and that is the part we correct at the top.
Today's Market Moves
August CPI MoM, 0.3% bucket
33.0%→33.0%0.0pp
THE TRADE THAT DIED. Ask 35.0. On assumed sd 0.05 our model said 40.7 and the edge was +5.7pp. On the measured 0.164 the model says 22.4 and the edge is -12.6pp. Same market, same day, opposite conclusion, and the only thing that changed was that we measured an input instead of guessing it.
August CPI MoM, >=0.5% bucket
13.5%→13.5%0.0pp
The measured model calls this +15.2pp CHEAP at the 14.0 ask. We are not taking that either, and the reason matters: at sd 0.164 the model is nearly flat and a flat model always says far buckets are underpriced. That is the barrier-model failure shape, arriving from the opposite direction.
Fed September 25bp increase
48.5%→51.5%+3.0pp
Bid 51 / ask 52 against a CME FedWatch near 49.4%. Back above a coin flip after Monday's holiday lull. We hold no expression on this for the first time since June, having closed both hike positions on Sep 7.
10Y Touches 4.8% Before 2027 (sold Sep 1)
93.7%→98.4%+4.7pp
Bid 97.3 / ask 99.5. Two basis points from resolving on a 4.78% Friday settle. Sold at a volume-weighted 71.69c into a ten-share book. $38.81 foregone at today's bid, $42.91 at resolution. Correct process, bad outcome, printed above everything that works.
Fed Rate End 2026 = 4.0% (pos-011)
41.2%→42.6%+1.4pp
Bid 40.6 / ask 44.6. FV recomputed at 40.6, published 41, unchanged. The ladder-versus-standalone cross-check we ran daily all last week is GONE - we closed the standalone contract yesterday, so this fair value now has no second reading. A real loss of information and a direct cost of yesterday's consistency argument that we did not price at the time.
US Unemployment 2026 >= 5.0% (pos-004)
9.75%→9.75%0.0pp
Flat, and still 2.20 above where it sat before a jobs report that was the best news this position will ever get. Fifth session without a verified explanation. Still 3.8 points cheap on our own fair value of 6, which is the only reason it survived a session that closed two others.
Screening Table
| # | Market | Expiry | Market Price | Fair Value | Gap (pp) | Direction | Volume | Confidence |
|---|---|---|---|---|---|---|---|---|
| 1 | August CPI MoM 0.3% bucket - RETIRED, not traded | Sep 11 | 33.0% | 0% | 0pp | NO TRADE - measured dispersion reverses the sign from +5.7pp to -12.6pp; category retired to diagnostic-only per Sunday's pre-commitment | $$1.1K liquidity on the leg | 5/10 |
| 2 | US Unemployment 2026 >= 5.0% | Dec 31 | 9.75% | 6% | +3.75pp | HOLD $100 NO - 3.8pp cheap at the size-aware bid; fifth session rising against good news with no explanation | $$124.6K on the leg | 4/10 |
| 3 | Fed Rate End 2026 = 4.0% | Dec 31 | 42.6% | 41% | -1.6pp | HOLD $25 YES - 1.6pp rich at the mid, 0.4pp cheap at the bid, and the cross-check that validated this FV is gone | $$1.37M on the leg | 2/10 |
| 4 | August CPI MoM >=0.5% bucket | Sep 11 | 13.5% | 0% | 0pp | NO TRADE - measured model calls it +15.2pp cheap, which is exactly what a flat model always says about tails | $$10.4K liquidity | 5/10 |
| 5 | Fed September 25bp increase | Sep 16 | 51.5% | 50% | -1.5pp | NO POSITION - back above a coin flip; no gap worth the name and the book deliberately exited this theme yesterday | $$16M on the leg | 3/10 |
| 6 | CONSTRAINT SWEEP - 56 exclusive ladders scanned | standing | 0% | 0% | 0pp | NO TRADE - the two real structural edges (1.4pp Fed Oct, 4.2pp China) are both smaller than the spreads and slippage needed to collect them | $$9.2K to $864K quoted depth | 5/10 |
Market vs Fundamentals
Market Price (red) vs Estimated Fair Value (green) — %
Top 5 Opportunities
1
We measured our own confidence and it was wrong by a factor of three — NO TRADE
↑ BUY YES0pp
Market price
33.0%
Fair value
0%
Gap: 0pp
Sunday we wrote down what today's backtest would mean before we could know what it would say. That was the whole point. The August CPI month-on-month 0.3% bucket screened at +5.6pp against the ask on an assumed standard deviation of 0.05. The Cleveland Fed publishes an evaluation of its own nowcasting model - Economic Commentary 2023-06, Knotek and Zaman, Table 1, real-time data September 2000 through December 2022 - and the RMSE for headline CPI at the last-day-of-target-month horizon is 0.164 percentage points. Our nowcast is stamped Sep 4 for August with release Sep 11, past the last day of the target month, so 0.164 applies. For a single-month nowcast the other eleven months are already known, so the year-over-year error is essentially the month-on-month error; that step is our inference and we flag it as one. Three and a third times too confident. On the measured number the bucket is 12.6 points rich rather than 5.6 cheap. We are not banking the reversal either, because the measured figure is unconditional across a sample that includes the 2021-22 inflation surge, and at 0.164 our model flattens to 10.0 / 15.1 / 22.4 / 23.3 / 29.2 against a market peaked at 2.6 / 6.0 / 33.0 / 48.5 / 13.5. A flat model always calls the tails cheap. We have seen that film.
▵ Bull case
- The measurement exists, is public, comes from the institution that builds the nowcast, and covers 22 years of real-time data. That is a far better input than anything we had yesterday.
- It arrived before the trade rather than after, which is the entire purpose of writing the pre-commitment on Sunday.
- The nowcast itself is good - it beats the Blue Chip consensus and the SPF across both long and pandemic samples. What failed was our dispersion assumption, not the anchor.
- Retiring the category costs us nothing today, because on every plausible dispersion the trade is somewhere between marginal and badly wrong.
▿ Bear case
- This category produced pos-006, pos-007 and pos-008 - three of the book's biggest early winners. We are retiring the thing that worked best.
- The pre-commitment may have been badly drafted. It anticipated a useless nowcast; what we actually found is a good nowcast and a bad assumption of ours, which arguably calls for fixing the assumption rather than retiring the category.
- We could not complete the bias check we promised - the paper reports RMSE but not mean error - so half the test we committed to is simply unfulfilled.
- 0.164 is the wrong number for a calm month and we know it. Retiring on a figure we have already argued is too wide is not obviously better than trading on one that was too tight.
2
Monday was Labor Day and yesterday's letter did not notice — CORRECTION
↑ BUY YES0pp
Market price
0%
Fair value
0%
Gap: 0pp
Yesterday we published a Daily Macro US letter dated Sep 7, written as a normal pre-market weekday edition. We screened markets, closed two positions, ran a gate check and discussed the session. Monday September 7 was Labor Day and the US market was shut. We found it this morning the same way a reader would: the Treasury par yield curve has no Sep 7 row. It runs Sep 1, Sep 2, Sep 3, Sep 4 and then stops. Two things follow and we want both on the record. The trades STAND. pos-013 and pos-010 were closed on Polymarket, which runs continuously, at bids we published with depth behind them, and a US holiday does not make those fills wrong. But the letter described a trading day that did not exist, and evaluated price-triggered gates against it one day after adopting a rule that those gates settle on the daily close. There was no daily close. One flag comes off as a consequence: we had been carrying the Cleveland nowcast's 09/04 stamp as a possible data-feed problem for three sessions. It was not a problem. The Cleveland Fed updates each business day, and Monday was not one. The carry was correct and our flag was wrong.
▵ Bull case
- We found it ourselves, within a day, from the primary source, and published it at the top rather than in a footnote.
- The trades are unaffected - continuous markets, published bids, verified depth.
- It resolves a stale flag we had been carrying against the Cleveland feed for three sessions, which was our error and not theirs.
- It is the second time in two weeks that checking the Treasury curve against our own narrative caught something, which argues for keeping that check.
▿ Bear case
- We adopted a daily-settle gate rule on Sep 7 and immediately applied it to a day with no settle. The rule was one letter old when we misapplied it.
- A publication that covers US macro every weekday should know the US market calendar without deriving it from a missing table row.
- It raises a question we have not answered: how many other gate checks have been run against sessions that did not exist? We have not audited that and are not pretending we have.
- Thin holiday liquidity is exactly when a size-aware bid mark is least reliable, and we marked the book against it without comment.
3
US Unemployment 2026 - fifth session against us on the best news it will get — NO
↑ BUY YES+3.75pp
Market price
9.75%
Fair value
6%
Gap: +3.75pp
Flat on the day and still 2.20 above where it traded before Friday's employment report - a report that printed payrolls at +162,000 against a 53-60,000 consensus, revised July from -23,000 to +23,000, and held the unemployment rate at 4.1% when consensus had it ticking to 4.2%. Our side should have fallen. It has gone 7.55 to 9.25 to 9.75 and stayed. This is the fifth letter in which we have written that we do not have an explanation, and we are still not going to manufacture one from the three candidates we keep listing - thin liquidity, a hiking Fed lifting the 2027 policy-error tail, internals narrower than the headline. None is verified. What keeps the position is arithmetic, not conviction: fair value 6 on the YES makes the NO worth 94 against a 90.2 size-aware bid, so we are 3.8 points cheap on the identical test that closed pos-013 at one point rich and pos-010 at 5.5 yesterday. Getting 4.1% to 5.0% by December still requires roughly 1.4 million additional unemployed Americans in under four months.
▵ Bull case
- Payrolls +162,000 with July revised positive and unemployment holding at 4.1%. The path to 5.0% got materially harder on Friday.
- 3.8 points cheap at the size-aware bid on the same test we used to close two positions yesterday.
- The hard auto-close at YES below 5c is written, untriggered and carries no discretion, so the downside process is already fixed.
- GDPNow at 4.7% and ISM manufacturing at 54.6 describe an economy that does not shed 1.4 million jobs in a quarter.
▿ Bear case
- Five consecutive sessions of the market moving against us on unambiguously good news, and we cannot explain it. That is the strongest bear case available and it is ours.
- When a disagreement persists and widens, the base case should shift toward the market being right and our fair value of 6 being wrong.
- Our 133.3333 shares walk two levels to clear, so the published mark flatters a hurried exit.
- We have just retired one model for overconfident dispersion. A fair value of 6 against a 9.75 market is its own confidence claim and it deserves the same scepticism.
4
Fed Rate End 2026 = 4.0% - we deleted our own cross-check yesterday — YES
↑ BUY YES-1.6pp
Market price
42.6%
Fair value
41%
Gap: -1.6pp
Up 1.40 on the day. Fair value recomputes to 40.6 and we publish 41, unchanged. The method has not moved: the end-2026 ladder prices 3.5% at 7.25, 3.75% at 22.85, 4.0% at 42.60, 4.25% at 21.35 and 4.5%-or-more at 5.10, a hike mass of 69.05; a September hike blends to 51.2 from a 51.5c market and a 49.4% FedWatch; 0.512 plus 0.488 times 0.30 gives 65.8 for a hike at some point in 2026; and 65.8 times 42.60 over 69.05 is 40.6. What HAS changed is that we can no longer check it. All last week we published the disagreement between this ladder's hike mass and the standalone hike contract - 9.3 points, then 9.0, then 5.4 - as a running diagnostic on whether the derivation could be trusted. We closed the standalone contract yesterday on a consistency argument, and in doing so deleted the second reading that made this fair value auditable. That is a genuine cost of yesterday's decision which we did not price when we made it, and it is why confidence on this position drops rather than rises even as the numbers converge.
▵ Bull case
- One hike and stop is the modal path on a 49.4% FedWatch, and this bucket pays $73.53 on a $25 stake if it lands.
- At the 40.6 size-aware bid we are 0.4 points cheap, so the position is not rich on the test that governs this book.
- The ladder shifted toward us on the day: 4.0% went 41.20 to 42.60 and 4.25% 19.05 to 21.35.
- $25 of stake at a 34c entry retains real asymmetry even after the spread.
▿ Bear case
- We deleted the cross-check that validated this fair value and we did it ourselves, one day ago, for reasons that had nothing to do with this position.
- At the 42.60 mid we are 1.6 points rich; the entire disagreement sits inside a 4-point spread either way.
- This is now the book's only rates expression, derived from a structure we hold no other reading on.
- Two hikes or none and the bucket pays nothing. It remains the narrowest possible expression of a view the book has otherwise exited.
5
The 4.8% touch is two basis points away and we sold it a week ago — YES
↑ BUY YES0pp
Market price
98.4%
Fair value
0%
Gap: 0pp
The contract trades 98.40 with a 97.3 bid, against the volume-weighted 71.69c we took on Sep 1 when the published 66c-bid take-profit fired into a book with ten shares on top of it. The ten-year settled 4.78% on Friday, two basis points from the 4.80% barrier. Our 151.5152 shares would fetch $147.42 today and $151.52 at resolution, against the $108.61 we collected. That is $38.81 foregone now and $42.91 if it prints. We keep this at the top of the letter for a reason that has nothing to do with self-flagellation: a book that reports its gates only in the weeks they pay has no process to describe, and this is the eighth consecutive session we have published the running cost of one that did not. The verdict is unchanged and will stay unchanged. The rule was written when this position's fair value was 60, it was published every day for a fortnight, and it was honoured on the morning it fired. What we conceded on Sep 2 stands too: a fixed take-profit that never tracks fair value sells winners early by construction, and ours sat at 66c while our own arithmetic went to 76.8 and beyond.
▵ Bull case
- The exit was the published rule working exactly as written, honoured on the day it fired without hindsight.
- Tuesday's book had ten shares at 82c. Today's 97.3 bid was not a choice available to us that morning.
- +$83.61 on a $25 stake is still the second-largest winner this book has produced.
- Closing it removed the largest correlated rates bet at a moment when the book has since exited the theme entirely.
▿ Bear case
- $38.81 foregone at today's bid and $42.91 at resolution, on a $25 stake. No framing shrinks those numbers.
- We knew fixed take-profits sell winners early and left ours at 66c anyway while fair value ran away from it.
- Our published fair value was frozen at 61 by a rule about a different market, which is what let the gate look reasonable when it was already stale.
- Eight sessions of publishing the same running cost is not the same as having fixed the gate-setting method, and we have not fixed it.