The light went back to yellow, and one billion dollars decided it

Where we are

The market-risk headline reads YELLOW today. It read RED on every single recorded day from September 1 through September 6 — six consecutive daily snapshots, no gaps. Last week's edition reported that turn and spent five sections on it. Today the light is back where it was before, and nothing in the world that the phrase 'risk is lower' would normally describe has happened.

What actually changed is one number's sign. The headline is a gate with two switches, and it is published rule rather than judgement: the light turns RED only when consumer sentiment is below 55 AND global central-bank money is classified as draining. Either switch alone leaves it YELLOW. The first switch is still on and cannot turn off this week — the University of Michigan survey printed 51.7 on August 28, it is below 55, and there is no new print until the end of this month. So the entire colour of this dashboard now rests on the second switch, alone.

That second switch turned off. Global central-bank money was reclassified from draining back to neutral on today's read, and the colour followed mechanically. The next section is about how thin the thing that moved it is, because the honest description of this week is not that conditions improved. It is that a measurement crossed back over a line it had crossed six days earlier, and the distance it travelled to do so is smaller than almost anyone would guess.

One billion dollars, out of twenty-five trillion

The rule behind the second switch is public and simple. Central-bank money counts as draining only when it has fallen over the last four weeks AND fallen over the last twelve weeks. Both must be negative. Either one positive and the label is neutral.

The twelve-week number is clearly negative: down $172B. The four-week number came out at plus one billion dollars.

The total it is measured against is $25.09 trillion. One billion on twenty-five thousand and ninety billion is 0.004% — four thousandths of one percent. That is not a magnitude, it is a sign bit. Had the same four weeks produced minus one billion instead of plus one, every public surface Jawz runs would be showing a red light this morning, and this edition would have a different title.

It is worth being precise about what that does and does not mean. It does not mean the drain was imaginary — the twelve-week fall is real and I will come back to it. It means the specific instrument that converts a continuous quantity into a colour has no dead zone around its threshold. The cycle quadrant on this dashboard does have one: it requires three consecutive daily reads agreeing before it will announce a change, which is why it has sat at SUMMER for twenty days without flickering. The liquidity switch has nothing equivalent. It flips the moment the sign flips, and so the headline can round-trip inside a week on a rounding error.

I said I would say this in these words

Last week's edition closed its watch list with this, as the first item:

“Whether the liquidity classification holds. It flipped to draining on a week whose step was partly the dollar; a week of dollar weakness could flip it back to neutral and turn the headline YELLOW again with nothing real having changed. If that happens, this brief will say so in those words, because a colour that follows the exchange rate is not a colour worth acting on.”

That is what happened, within the week, and this is the brief saying so in those words. Over the same four weeks that produced the deciding plus-one-billion, the dollar fell 0.31%. A weaker dollar mechanically inflates every non-US central bank's balance sheet when it is translated into dollars — the European, Japanese and Chinese legs all get bigger without a single asset being bought. That is the direction that pushes a four-week change from negative toward positive, and it is the direction the dollar moved.

I am not claiming the currency accounts for the entire billion; on a margin this thin, nothing accounts for the entire billion, which is rather the point. I am claiming what I said I would claim if it happened: the colour changed, the world did not, and a reader who treats this morning's yellow light as news about risk has been misled by an instrument I built.

The deciding switch runs on a number that has not changed since July

There is a sharper version of this problem, and it took looking at the code to see it.

This brief has always reported the liquidity trend on a constant three-bank basis — the Fed, the European Central Bank and the Bank of Japan, measured the same way at both ends — precisely so that a bank entering or leaving the count cannot fake a move. That is the right way to report a trend, and it is what the tables show.

But the classification that decides the colour is not computed on that basis. It is computed on the four-bank total, China included. And China's contribution is $7.46T derived from a single figure — ¥50.207T — dated July 1 and published on August 15. It has not changed in 68 days and will not change until Beijing publishes again. In dollars, therefore, it moves only when the exchange rate moves.

Here is what that is worth. A move of 0.0134% in the yuan — from 6.7260 to 6.7269 per dollar, a shift in the fourth decimal place — is worth one billion dollars on that frozen leg. One billion dollars is the entire margin that decided the colour. And the exchange rate doing that work is itself a week-old quote, last observed August 28.

So the deciding switch of the headline can be moved by the fourth decimal place of a stale currency quote applied to a two-month-old Chinese accounting statement. The trend underneath is a different and better-founded thing: on the constant three-bank basis, central-bank money fell from $17.84T to $17.63T over twelve weeks, down 1.18%, with the Fed shrinking steadily in its own currency throughout. The drain over the quarter is real. The precision of the line drawn across it is not.

What did not move

Almost everything else, which is the useful context for a colour change this thin.

The business cycle stays in SUMMER — the quadrant where the economy grows and prices rise together — confirmed on twenty consecutive daily reads with high confidence, not sitting on its boundary, unchanged for a third straight edition. Inflation is still the dominant risk under the read, with core PCE, core CPI and the market's five-year inflation forecast all rising.

Financial conditions read loose and easing. Riskier companies pay 2.65% extra to borrow, which is 0.10 points less than four weeks ago. Safer companies pay 0.81%. Expected stock-market volatility is 14.3, the lowest reading in the twelve-week window. Credit is not stressed and volatility is not elevated; by the measures taken from market prices rather than from surveys and classifications, this is a calm week.

The one pillar that keeps moving in the same direction is the ten-year government bond yield after subtracting expected inflation — the real cost of borrowing — at 2.42%, classified restrictive, up 0.27 points over twelve weeks. That is the fourth consecutive edition naming it as the only persistent mover, and it is measured in market prices, which is why it deserves more weight than the colour above it.

So the standing tension is unchanged and, if anything, better lit than last week. Market prices say calm. Households say worse. Central banks over the quarter have on balance removed money. Those three have disagreed for six editions now, and this week's headline flip contributed no information about which of them is early.

Six days of red have disappeared from the record

One more disclosure, of the same kind as last week's and for the same reason.

Before this morning's data refresh, two Jawz tools disagreed about today. The live engine read YELLOW. The regime history tool, sampled weekly, showed today's row as RED and listed a transition — 'September 7: YELLOW to RED' — on a day the engine itself read YELLOW. Anyone connecting to Jawz in that window and asking what changed could have been handed that date as a fact.

After the refresh it is worse, and in the opposite direction. The same tool, same twelve-week window, sampled the same way, now reports the market regime as YELLOW continuously since August 3, with no colour change since. The six consecutive days of RED from September 1 to September 6 — the entire subject of last week's edition, recorded live in the daily table, still sitting there — appear nowhere in it. The list of transitions went from five to four. An entire regime episode vanished from the history, and the tool reports the result with no gap and no warning.

The mechanism is ordinary and that is the problem. The weekly view samples one day in seven: it looks at August 31, which was YELLOW, and at September 7, which is YELLOW, and correctly concludes that nothing changed between two dates it actually looked at. The RED episode began and ended in the six days it never sampled. Add the second mechanic — each sample date is matched to the nearest stored daily reading within three days — and the same tool can also show a stale row wearing today's date, which is exactly what it did at 5am. Neither behaviour is a coding error. Both are the sampling grid deciding what counts as history.

This also means last week's edition and this one disagree with the tool and with each other on a date. That brief dated the turn to 'the September 2 read'; the daily table says the first RED day was September 1. I am not going to quietly change a published sentence to match — corrections here go through Pedro and get printed with their reason, the way the August 3 one did. It is flagged, and it will be corrected in the open if confirmed.

Why this earns a section rather than a footnote: the standing rule of this brief is never to announce a shift without a transition date, and the tool that supplies those dates has now been wrong in both available directions within six hours — a transition invented on a day it did not happen, and six days of transitions erased as though they never did. A reader of the weekly view this morning would conclude the light has been steady for five weeks. It has not. Last week the failure was a warning that went quiet because the risk moved to a variable it never watched. This week it is a history that is smooth because it is not looking often enough to be bumpy. Both are the same species: an instrument answering the question it was built to answer, while the reader believes it answered a different one. If you take one thing from this edition, take that a smooth history is sometimes a claim about the sampling rate and not about the world.

What to watch

• Whether the colour flips again, and how fast. The four-week liquidity change is sitting within a billion dollars of zero on a twenty-five-trillion base, with no dead zone around the line. Another round-trip in the next fortnight would not be a surprise, and would not be news about risk either. If it happens, this brief will keep saying so.
• The Chinese statement, now 68 days old and still the July figure. The August figure is due around the middle of this month, and the release watch confirms nothing newer has been posted. Until it lands, the fourth leg of the measure that sets the headline colour is a constant that moves only on the exchange rate.
• The consumer survey's next print at the end of this month. It decides two separate things: the level, which has held the first RED switch on since August 28, and the three-month trend, which is still positive and still one of three components holding the cycle in SUMMER. A second consecutive fall would put both on the same side for the first time.
• The real cost of borrowing at 2.42%, restrictive and up 0.27 points over twelve weeks while every other pillar eased. It is the one measure that has moved the same direction all quarter, and it is priced by markets rather than classified by a threshold.
• Data quality is clean this week: every source current, nothing stale or unavailable, the collector healthy. Eleven of the thirteen weekly rows behind this edition were recorded live; the two at the old end are reconstructed from point-in-time data — each series as it was published on that date, not as later revised. One defect stays open and undiagnosed for a fourth week: the August 17 edition described that week's conditions direction as easing, and the stored row still reads stable with the change dated August 24. Unchanged from the last three disclosures.

The frame, not the trade

The specific way to be wrong this week is to read the yellow light as an improvement and take some risk back. Nothing in this week's data supports that. The gate reopened because a four-week arithmetic came out at plus one billion instead of minus one, on a basis whose fourth component has not been updated since July. Six days of red preceded it and the quarter-long drain underneath is unchanged.

The generalisable lesson is about thresholds, and it is the sequel to last week's. Any rule that converts a continuous quantity into a discrete state — a colour, an alert, a stop level, a rebalancing band, a covenant — has a region near its boundary where the output is decided by noise rather than by the thing being measured. The output does not look any less confident there. A red light and a yellow light render identically whether the underlying margin was forty percent or four thousandths of one percent.

So the question for your own book is: which of your rules are currently sitting near their lines, and would you be able to tell? Take any threshold you actually act on and ask how far the input is from flipping it, in the same units the input moves in day to day. If the answer is 'less than it moves in a normal week', then that rule is not currently giving you a signal — it is giving you a coin flip with a confident colour on it, and the right response is to widen the band, add a confirmation delay, or stop treating its output as information until the margin is real again.

None of this is advice and none of it is a forecast. It is a week's data, the reasoning applied to it, a headline that changed for a reason too small to act on, and an instrument of mine that got a date wrong in public.