The six sessions

The taker buy share answers one question: of the volume that crossed the spread, what percentage was buyers hitting offers rather than sellers hitting bids. Below 50% means sellers were the side in a hurry.

SessionCloseChangeBuy shareVolume
Fri Sep 4$79,6611.98% lower48.36%$1,538M
Sat Sep 5$79,832+0.21%46.67%$727M
Sun Sep 6$80,342+0.64%47.58%$707M
Mon Sep 7$79,1121.53% lower49.00%$840M
Tue Sep 8$78,4560.83% lower44.42%$1,510M
Wed Sep 9, partialabout $78,810+0.45%46.29%$890M

Mean 47.06%, low of 44.42% on Tuesday, and price down 3.03% across the run from $81,270.

Look at the change column before going further. Three of the six sessions closed higher even though sellers took the larger share of aggressive volume in every one of them. That is not a contradiction, it is the whole point of the metric: it measures who was impatient, not where price ended up. We have written that sentence many times and this week is the cleanest example of it we have had.

Runs of six or more happen about 31 times in 1,000 sessions, so roughly monthly. This is the sixth of 2026, after January 15, February 19, May 31, June 15 and July 7. What stands out is the gap: the last one ended July 13, 58 days ago.

And positioning went the other way

DateLong/short account ratio
Sep 51.0165
Sep 61.0475
Sep 71.0602
Sep 81.1608
Sep 91.2952

Five consecutive increases while the tape was one sided the other way. Sellers were urgent every session, and accounts got steadily longer into it. Both statements are true of the same market, because they answer different questions: one is about who was in a hurry, the other about what people are left holding.

Test one: 80%, and completely wrong

The obvious question is whether a week of one sided selling says anything about the following week. We have 1,000 sessions of buy share data, so we asked it.

First attempt. Find every completed streak of six or more sub-50 sessions, take the last day of each, and measure the next five sessions. Thirty streaks qualified.

The result looked excellent. Twenty four of thirty higher five sessions later, 80%, mean +2.81%, against a baseline where only 52.5% of all sessions are higher five days on. That is the kind of number people build strategies around.

It is worthless, and one extra check is what exposed it. We ran the same test at a one session horizon and got 30 positive out of 30, a perfect record with a mean of +3.02%.

Nothing in markets is 30 for 30. When a backtest returns a perfect score, the answer is always that the test knows something it should not.

Here it did. A streak ends, by definition, on the first day the buy share climbs back above 50%. And the buy share correlates with that same session's return at about 0.5. So "the day after a streak ends" is not a neutral future date. It is a date we selected because buyers came back on it. We had defined an up day into the sample and then congratulated ourselves for finding one.

Every number in that first test inherits the same flaw, including the 80%.

Test two: 63.7%, and still wrong

Second attempt, with the look ahead removed. Ask a question that can actually be asked in real time: today, and the five sessions before it, were all below 50%. No information from tomorrow. This is exactly the situation we are in as we write.

That condition is met on 147 of 995 sessions, about 15% of the time.

HorizonSignal, meanSignal, positiveBaseline, meanBaseline, positive
1 session+0.16%51.4%+0.09%50.8%
5 sessions+0.62%63.7%+0.46%52.5%
10 sessions+1.38%59.6%+0.91%53.2%
20 sessions+0.74%49.3%+1.88%51.9%

The five session row is the seductive one: 63.7% against a 52.5% baseline. Eleven points better than chance, on what looks like 147 observations.

Except it is not 147 observations. Those 147 days cluster into only 31 distinct episodes. An eleven day streak contributes six separate "observations" whose forward windows overlap almost entirely, and they are not independent measurements of anything. They are one event, counted six times, which makes a small sample look like a large one.

Notice the last row too, which we would have been happy to skip. At 20 sessions the signal is worse than the baseline on both measures: +0.74% against +1.88%, and 49.3% positive against 51.9%. Whatever the effect is, it does not survive a month.

Test three: one observation per episode

The fix is simple. Count each episode once. Take the first day the condition is met, ignore the overlapping days inside the same streak, and measure forward from there.

Version of the testSampleMean, 5 sessionsPositive
One, with look ahead30 streak ends+2.81%80.0%
Two, overlapping days147 days+0.62%63.7%
Three, one per episode30 episodes0.69% lower56.7%
Baseline, all sessions995 days+0.46%52.5%

The mean goes negative. Same data, same condition, three answers, and the only one computed correctly says that after six sessions of net selling the average five session return is worse than doing nothing. The hit rate stays a little above baseline at 56.7%, and at ten sessions the mean recovers to +1.07% on 60% positive, so we would not claim the effect is exactly zero on 30 episodes.

What we will claim is this: there is nothing here anyone should trade. A result that falls from +2.81% to negative depending on how you count is not a finding, it is an artefact wearing a finding's clothes.

Why publish the failures

We could have run the third test first, printed one honest paragraph, and looked more competent. We are showing all three because the two failures are the transferable part.

Look ahead bias is rarely obvious. Nothing in test one used tomorrow's price. It used tomorrow's buy share, through the definition of when a streak ends, and that was enough to manufacture an 80% hit rate. If you read a backtest built on "the day after the signal ended", ask what defines the ending.

Overlapping windows inflate confidence, not just precision. 147 sounds like a sample that settles an argument. 31 does not. Any test on consecutive-day conditions has this problem by construction.

A perfect score is a bug report. The 30 out of 30 was the only reason we looked again. If test one had returned 73% we might have published it, and you would have had no way to know.

That last sentence is the uncomfortable one, and it is why this article exists.

Three true framings of the same price

Bitcoin trades near $78,810. Here is that number three ways, all correct.

FramingReferenceChange
Year to date2026 opened at $87,64810.12% lower
From the high$126,200 on Oct 6, 202537.6% lower
From the low$58,625 on Jun 30, 2026+34.4%

Nobody is lying when they pick one. The recovery is real and the year is still red, and which fact leads a headline depends entirely on what the writer wants you to feel. We would rather print all three and let the arithmetic sit there.

The 50-week moving average now stands at $80,373, so price is about $1,560 below it. Five days ago Bitcoin closed above that line for the first time in 296 days and held it for exactly one session.

What is coming, and what we cannot see

Two dated events sit in front of this. US inflation data on Thursday, September 10. Then the Federal Reserve on September 15 and 16, with the Senate's CLARITY Act cloture vote at 2:15pm Eastern on the 15th. Oil has added about 10% this month with WTI above $94, which is the kind of input that makes an inflation print less predictable.

And the largest thing our data cannot see is doing something interesting. Reporting puts August spot Bitcoin ETF inflows at $3.52 billion, including the biggest single day since January at $731 million, and yet net flows for 2026 remain roughly $1 billion negative. A record month inside a negative year. Those creations settle away from the public order book, so no buy share on any venue registers them, and the six sessions above are silent about all of it.

Which is the honest closing note. We measured a real pattern this week, tested it three ways, and found that the only correct test says it is not worth acting on. Meanwhile the flow that might actually matter is happening somewhere our instrument does not reach.