Yesterday this desk published an article about a three-pip disagreement. Three publishers put the dollar–yen’s two-hundred-day moving average at 158.50, 158.52 and 158.53; Tokyo’s high was 158.51, inside the band and equal to none of them. We said in print that we had not checked whether all three were simple averages and that this alone could account for the spread. We have now checked, and it does not account for the spread. It dwarfs it. A fourth publisher gives the two-hundred-day as an exponential average sitting just under 158.00 — about 52 pips below the middle of the band we were arguing about, seventeen times as far as the whole dispute. Tonight the pair closed at 158.04, which is four pips above one of those numbers and forty-eight below the other.
We published the wrong error bar by a factor of seventeen
The three readings in yesterday’s piece were 158.50 from MUFG via FXStreet at 10:07 UTC, 158.52 from Minkabu’s technical page at 07:40, and 158.53 from Gaitame’s morning outlook filed the night before. Three pips. We treated that as the precision available on the level, told you to place your stop on your own loss tolerance rather than on any of them, and moved on.
At 23:08 UTC tonight FXStreet filed a fourth reading, and it is the first one that names its convention: a two-hundred-day exponential moving average, described as sitting just under 158.00, with the additional claim that it has been below every daily close since 1 October. Against the 158.50 to 158.53 band that is a gap of 50 to 53 pips, 52.5 to the band’s midpoint.
So the honest version of yesterday’s article is that we measured the disagreement between three implementations of one convention and published it as the uncertainty on the indicator. The uncertainty on the indicator is the convention, and it is an order of magnitude larger. If you took our three-pip figure as the error bar on a level you were trading, we gave you a number that was wrong by a factor of seventeen, and we gave it to you with the arithmetic shown, which is the kind of wrong that is harder to notice.
The same publisher changed its own figure by a pip between breakfast and dinner
There is a smaller finding underneath, and it is the one that kills the comfortable explanation. Yesterday’s three-pip spread could have been three publishers using three different closing conventions on three different price feeds — irritating, but stable, and something you could pick one of and stay with.
Gaitame.com published three figures for its own two-hundred-day across thirty-six hours. Its outlook filed at 23:24 UTC on 6 October gave 158.53 for the coming session. Its technical table filed at 17:15 UTC on 7 October gave 158.52, with its own previous-session field also reading 158.52. Its outlook filed at 23:00 UTC tonight gives 158.52 for yesterday and 158.53 for today.
Read that as a single publisher’s view of 7 October and it is 158.53 in the morning, 158.52 by the evening, and 158.52 once the day is over. One pip, one publisher, one day, revised intraday and then settled at the lower value. That is a tenth of the convention gap and it is not worth a basis point of anybody’s risk. It is worth recording because it removes the option of solving the problem by choosing a source. Choosing a source leaves you with a figure that still moves while you hold it.
The same table gives the five-day average at 158.00 against 157.87 the session before, the twenty-one-day at 156.73 against 156.54, the ninety-day at 159.54 against 159.57, a fourteen-day RSI of 53.99 per cent, Bollinger bands at 159.76 and 154.02, and a parabolic stop at 155.81. Those are single-sourced and none of them is load-bearing here.
Tonight the two conventions say opposite things, and both of them say the average held
This is the part with money in it. Wednesday’s close was about 158.04 and Tokyo opened near 158.00.
On the exponential reading, the pair is four pips above its two-hundred-day, and the publisher carrying that reading says it has closed above it every day since 1 October. That is a pair in an uptrend that has respected its long average for a week.
On the simple readings, the pair is 48 to 49 pips below its two-hundred-day, and the single time it reached up to touch it — 158.51 yesterday, one pip above the lowest of the three — it was rejected and spent the rest of the session forty pips lower. That is a pair capped by its long average.
Both sentences are “the two-hundred-day is working.” They are statements about opposite things. A rule of the form long while above the two-hundred-day is long on one convention and flat or short on the other, on the same instrument, on the same evening, with no ambiguity in the price and none in the data. The ambiguity is entirely in a word.
So: if your system has a condition keyed to a named indicator, the condition is not specified until it carries three things — the convention, the publisher, and the time of day the value was read. This pair gives you 52 pips on the first, a pip on the second and third, and a one-pip difference in the fit to yesterday’s high, which is the only one of the three that anybody wrote about.
Three filters, four pips: a correlation you are probably not accounting for
A second consequence, and it is the one we would act on first. The five-day average is 158.00. The exponential two-hundred-day is just under 158.00. The close is 158.04. All three sit inside four pips of each other.
If your schedule contains a short-term trend condition and a long-term trend condition, and the short one is a five-day average and the long one is an exponential two-hundred-day, then tonight those are not two conditions. They are one condition sampled twice, and they will flip together on the same tick. The five-day moved thirteen pips in a single session, from 157.87 to 158.00, so the coincidence will not last — but the reason to notice it is not that it is durable. It is that nothing in either filter’s definition tells you when it is happening, and an exposure limit that assumes two independent confirmations will double your size at exactly the moment the two have collapsed into one.
That is a correlated-exposure problem wearing the clothes of a confirmation rule, and it is checkable in one line of code against whatever indicator values your platform computes. Which, on tonight’s evidence, may not be the same values anybody else is computing.
The window we specified held, and a calendar field was 46.6 basis points from any real number
Yesterday’s pre-US piece argued that a sixty-minute blackout either side of the 17:00 UTC ten-year reopening and the 18:00 UTC September FOMC minutes is one continuous window from roughly 16:00 to 19:00, not two, and that a schedule holding them as separate entries with independent pause logic would restart exposure for an interval that does not exist. Score it.
The auction cleared at 5.300 per cent on 39 billion dollars, a bid-to-cover of 2.77 described as the highest since 2016, with indirect bidders at 80.34 per cent and direct at 17.12 — which sums to 97.46 against a record non-dealer share reported elsewhere as 97.5, and against 100 minus the 2.54 per cent that went to primary dealers. The identity closes to four hundredths. The minutes reported all nineteen participants backing September’s quarter-point increase and most seeing another as appropriate by year-end.
The pair went from about 158.40 before the auction to about 157.90 after it, fifty pips, and closed at 158.04. The ten-year ran to 5.36 per cent at 13:25 UTC and ended near 5.28, an eight-basis-point round trip that finished within a fifth of a basis point of Tuesday’s close. And the full-day range was 158.51 to 157.85, 66.0 pips — exactly the figure we measured at 11:30 UTC, five and a half hours before the first event. The US session low of 157.90 sat five pips above London’s. Neither event produced a new extreme. The window was correctly specified and the thing it was protecting you from turned out to be a fifty-pip round trip inside a range that was already set.
One field we refused to publish yesterday is now settled. A vendor preview gave the previous yield on this auction as 4.834 per cent; we cut it as a benchmark and said in print why. It cleared at 5.300, which is 46.6 basis points away, and nothing in the ten-year complex traded near 4.834 at any point in the window. The field was not a stale yield from this security. It was not this security’s yield at all.
What this does not tell you
The exponential reading is one publisher, is given as “just under 158.00” rather than as a figure, and we did not verify its claim that the average has been below every daily close since 1 October. If the real value is 157.90 the gap to the simple band is 62 pips; if it is 157.99, it is 53. The argument does not depend on which, but the number does, and we are not going to print a two-decimal value that nobody published.
We also have not established that the three readings in yesterday’s band are all simple averages. None of the three publishers said so. We are now reasonably confident they are, because an exponential average on this data sits half a yen lower and all three are within three pips of each other — but that is an inference from the spacing, not a statement any source made, and it is exactly the kind of inference this article is about.
The session levels are secondary readings, with the close given as approximate by the publisher that reported it, and the day’s low of 157.85 comes from one publisher’s London report. The auction bidder shares come from one source and the non-dealer total from another; they reconcile, which is why we used both. We did not reach the Treasury’s own result page. And we still do not have an offering size for Thursday’s thirty-year from any issuer document — the arithmetic of a 119 billion dollar week less 58 for the three-year and 39 for the ten-year leaves 22, which is our subtraction and not a published figure.
Related
- FX events calendar — auction and minutes times in UTC, and the blackout windows they imply.
- Signals — what we require of a level before it carries a stop.
- EA presets — indicator conventions, filter independence and correlated exposure.
