The Side the Tape Leaves Out

An off-exchange trade is hidden while it works and public the moment it is done. Within ten seconds the tape carries its price, its size and the fact that it happened away from an exchange. The one field it never carries is the side: nobody on the tape is marked as the buyer. The rule the folklore uses to fill that field in — a print above the volume-weighted average price is accumulation, a print below it is distribution — fills it in with the day’s price path instead, because a running average is made of the prices behind it. On a steadily rising session that average sits half a session of drift below the current price, so every print is above it before anybody has bought anything. Cut the curriculum’s sixty closes into six ten-bar sessions and the rule’s accumulation share tracks the session return at a correlation of 0.93, which puts 87 per cent of the variation in one down to the other, and the customary 70 per cent filter fires on all three rising sessions and none of the three falling ones. It is not reading intent. It is reading the chart you already had.

Prerequisites: Lesson 48, which put the mean lag of a ten-bar average at 4.50 bars and is where the arithmetic below comes from, lesson 53, for the midpoint and why a fill that crosses there tells you nothing about who wanted it more, and lesson 54, for where your own order goes when it never reaches an exchange, which turns out to be most of what “the dark” is.

Six sessions, and what the rule actually measures

Cut the sixty closes into six consecutive stretches of ten and treat each as a session. For each one, compute the return from first bar to last, then recompute the running average bar by bar and count how many of the nine prints after the first sit above it. That count is what the folklore calls the accumulation share.

Session	Open	Close	Return	Prints above the average
Bars 1–10	100.7	103.7	+2.98%	8 of 9, or 88.9%
Bars 11–20	106.7	101.6	-4.78%	2 of 9, or 22.2%
Bars 21–30	100.5	100.0	-0.50%	3 of 9, or 33.3%
Bars 31–40	99.5	101.3	+1.81%	8 of 9, or 88.9%
Bars 41–50	101.7	105.9	+4.13%	9 of 9, or 100.0%
Bars 51–60	106.7	106.5	-0.19%	5 of 9, or 55.6%

The two right-hand columns are the same column. Their correlation is 0.93, so 87 per cent of the variation in the accumulation share is accounted for by the session return alone, on a series where nobody accumulated anything because there are no institutions in it. And the customary filter behaves exactly as that correlation predicts: 70 per cent above the average is cleared by sessions one, four and five, which are the three that rose, and by none of sessions two, three and six, which are the three that fell. Three for three in both directions.

The one session that is worth staring at is the last. It returned −0.19 per cent, which is nothing, and the rule split it five to four, which is also nothing. That is the pattern in miniature. When the day has a direction the rule reports the direction; when the day has no direction the rule reports no direction. It never reports anything the last hour of the chart did not already show you, and it costs more to compute.

Six sessions is six observations, and lesson 19 would refuse to settle much on that. It does not have to. The unanimity result is arithmetic and needs no sample at all: a monotone path puts every print on one side of its own running average, always, for every instrument and every session length. The six sessions are here only to show that the arithmetic survives ordinary noise, and 0.93 says it survives it comfortably.

What this does not settle

That the sixty closes are a session. They are the daily series this course has carried since lesson 38, read here as though each close were an intraday bar. Nothing in the argument depends on the horizon — the running-average lag is scale-free — but real five-minute bars are noisier relative to their drift than daily closes are, which would push the six accumulation shares closer to half and the correlation below 0.93. The direction of that error is knowable and its size is not, so treat 0.93 as the figure this series gives rather than the figure your instrument will.

That equal volume per bar is a fair stand-in for a real volume-weighted average. It is not: intraday volume is heaviest at the open and the close, so a real running average is pinned harder to the opening prices for most of the day and drifts up to meet the price only at the end. That makes the lag larger than the arithmetic above, not smaller. The concession runs in favour of the finding, which is worth saying out loud rather than quietly banking.

That everyone states the rule this way. Some versions compare the print to the contemporaneous quote midpoint instead of to the session average, and that version is a real signing rule with a literature behind it and a known error rate. It is also unavailable here, because an off-exchange fill usually happens exactly at the midpoint, which is the one case the signing rule cannot resolve. The folklore did not choose the session average over the midpoint out of carelessness; it chose it because the better reference returns nothing on exactly these trades.

That the volume figures are current. Around 45 per cent off-exchange and roughly a third of that in alternative trading systems are the shape of recent years, not a measurement of this month, and both move. The arithmetic they feed is what matters and it is robust to a fair amount of error in both: at 40 per cent off-exchange and a third in dark pools it is 13 and 27 rather than 15 and 30, and internalisation is still the larger half.

That the reporting deadline is exactly ten seconds. It is the figure in FINRA’s rules and the one every account of this quotes, and reporting deadlines get revised. Nothing above turns on it. If the deadline were one second the print would still arrive without a side, and if it were a minute the pattern read over three days would be unaffected.

And the concession that costs this lesson most: refuting one rule is not refuting the field. There are people who build off-exchange statistics from the same public data and a great deal of modelling — separating auction volume, netting the internalised retail flow out, conditioning on the quote at the moment of the print — and this lesson has not tested any of them. What it has shown is that the version you will meet in a forum post, a screenshot or a paid alert reduces to the day’s return. That is the version worth being able to refuse. Whether a careful construction on the same data carries information is an open question, and one this page has not earned the right to answer either way. And having spent a whole page taking a signal away, this lesson has not put one back, which is a real cost to the reader and not a rhetorical flourish.

Problems

1. Run the rule against the chart on your own instrument. Take twenty sessions. For each, compute the session return and the fraction of five-minute bars closing above the running volume-weighted average, which is one column of a spreadsheet. Plot one against the other and take the correlation. If it comes back near 0.9 you have reproduced this lesson; if it comes back near zero on your instrument, that is a genuine finding and worth writing down. An hour, and it settles the question for the market you actually trade.
2. Find out what your feed is really telling you. Take one off-exchange print on your platform and write down every field it shows. Does it name a venue, or does it show a single off-exchange flag? Then pull FINRA’s ATS transparency file for the same symbol and the same week, when it publishes two or four weeks later, and see whether the per-venue totals could have been known on the day. Twenty minutes, and it tells you whether the venue read you have been given is a reading or a label.
3. Measure how much of your instrument is actually blocks. For one week, record every print over ten thousand shares and total them, then divide by the day’s consolidated volume. In most names the answer is a few per cent, which is the honest size of the thing the whole story is about. Compare it with the share you assumed before counting. An evening, and the gap between the two numbers is the part of this that was never evidence.

Sources. FINRA Rules 6380A and 6380B, for the trade-reporting obligation and the ten-second deadline that makes the print public while the order stays private. FINRA ATS Transparency Data, for per-venue off-exchange volume and for the two-week and four-week publication lags that decide what a live feed can honestly claim to know. Charles Lee and Mark Ready, “Inferring Trade Direction from Intraday Data” (The Journal of Finance, 1991), for the standard way of signing a trade from its price and the quote, and for why a fill at the midpoint is the case it cannot resolve. Haoxiang Zhu, “Do Dark Pools Harm Price Discovery?” (The Review of Financial Studies, 2014), for what off-exchange trading does to the price on the lit book, which is the question worth asking once the intent-reading question is closed.

The order is hidden and the print is public, and the field that matters is neither: it is the side, and the tape has never carried it. The rule that reconstructs the side from the session average reconstructs the session instead. It calls every print on a smoothly rising day accumulation before anyone has bought anything, it tracks the six sessions cut from that series at a correlation of 0.93, and its 70 per cent filter fires on the three that rose and none of the three that fell. Reverse the buyer and the seller in every one of those trades and the tape is identical, which is the whole argument in one sentence. What the print does tell you is real and much smaller: that size traded, at that price, at that moment, away from an exchange. Lesson 57 takes that seriously and follows the order itself — what an institution actually does with a million shares, over how long, and what the doing of it costs.

Educational only. Trading involves substantial risk of loss. Not financial advice. Past performance does not guarantee future results.