The GLS Refresh Stamp Hides Four Different Data Frequencies

The composite on the Global Liquidity Scorecard read 85 when I captured it, and the refresh stamp beside it said 08:05 AM. Those two figures sit on the same header row and the eye takes them as one statement. They are answers to different questions. The 85 is a score. The 08:05 is the moment the score was last recomputed. Neither is the as-of date of the data underneath, and for a number that ends up in a client report, a committee pack or an attribution note, the as-of date is the only one of the three that carries any obligation.

This is not a criticism of the panel. Every macro composite ever built has the same property, because the world publishes its inputs on incompatible calendars and a composite has to print something. The problem is that a single timestamp implies a single vintage, and there is no single vintage. There are at least four.

What the recompute clock is actually measuring

GLS aggregates eight central banks, the Fed, the ECB, the Bank of Japan, the People's Bank of China, the Bank of England, the Swiss National Bank, the Bank of Canada and the Reserve Bank of Australia, and its stated inputs are the aggregate central bank balance sheet, global M2 money supply, USD liquidity indicators, credit spreads and the composite score itself. Five input families, eight jurisdictions, one number on a 0 to 100 scale.

The refresh stamp is a statement about the pipeline. It says the job ran at 08:05, pulled whatever each source was serving at that moment, and recalculated. If nothing new had published since the previous run, the composite would still stamp 08:05 and would still be showing you a figure whose slowest leg had not moved in three weeks. A refresh stamp that updates on a schedule tells you the plumbing is alive. It tells you nothing about whether the information is new.

The four clocks inside one composite

Sort the input families by how often the underlying data can change and the structure falls out immediately.

The eight central banks do not even agree with each other. Publication calendars range from weekly statements to monthly accounts, released on each institution's own schedule and in each institution's own currency. There is no hour of any day on which all eight are simultaneously current. An aggregate balance sheet figure is therefore always a mosaic, with some tiles from last Thursday and some from the close of a month that has already been superseded by the one you are living in.

Deriving the effective as-of date

The rule is simple and it is the one thing I would want a junior analyst to internalise about composites. The effective as-of date of a blended indicator is the date of its oldest binding input, not the date of its newest, and not the date it was computed.

Binding is doing work in that sentence. An input is binding if it carries enough weight to change the reading. A daily credit spread that moves the composite by a fraction of a point is not what your as-of date should key off. A monthly money supply series that is one of five named input families almost certainly is.

So the derivation goes like this. Take the slowest heavily weighted family, which here is money supply. Add its reference period, meaning the month it describes, to its publication lag, meaning the weeks between month end and release. That sum is roughly how far back the composite's slowest leg reaches. On a monthly series published a few weeks after the month it covers, a composite recomputed this morning is carrying a money supply leg that describes conditions from somewhere between one and two months ago. The stamp says 08:05 today. The information content on that leg is from a different quarter of the calendar.

Write both dates down. In practice I carry two fields next to any liquidity reading I cite. One is the recompute time, because it establishes that I did not quote a screenshot from last week. The other is the effective as-of, derived as above, because it establishes what was actually knowable.

The step that is a data event, not a liquidity event

Mixed frequency composites have a specific pathology that shows up in attribution. When a monthly series finally publishes, the composite can move meaningfully in a single recompute, with no market having done anything at all. The score steps because the pipeline swapped a stale month for a fresh one.

If you are running any process that reacts to changes in the composite, and particularly if you are running one that logs why a position was adjusted, that step will be recorded as a liquidity shift. It was a publication. The two look identical on the panel and they mean completely different things, because a market move is information the market already had and a release is information arriving.

The defence is to keep a release calendar alongside the composite history and mark which recomputes coincided with a scheduled publication. Any move on a publication day gets flagged before it is attributed to anything. It also fixes a subtler problem, which is that mixed frequency smoothing makes a composite look calmer than the world it describes. A monthly leg holds its value flat for weeks and then jumps. Averaged into a score, that produces long plateaus and sudden discontinuities, and a reader who has not been told about the plumbing will read the plateau as stability.

What to put in the pack

The version I use in committee documents has three lines and takes one row of a table.

1. The reading and its scale, stated exactly as the panel gives it. Composite 85 on a 0 to 100 scale, regime RISK-ON, policy EASING. No paraphrase, because paraphrase is where a score quietly becomes a forecast.
2. The recompute time, with the date. 08:05 on the capture date, not "this morning", which stops meaning anything the moment the document is filed.
3. The effective as-of, with one clause saying which input drove it. Something of the form: slowest weighted input is monthly money supply, so the reading is effectively current only through the most recently published reference month.

The reason this is worth the thirty seconds is the review that happens later, when a position taken in a RISK-ON read has gone against you and somebody asks what you knew on the day. If the answer is a screenshot with an 08:05 stamp on it, you have documented that you looked. If the answer includes the effective as-of, you have documented what you looked at, which is the question actually being asked. Those are different defences and only the second one survives the follow-up.