.webp)
A trader's observation about the 200-day average was tested against the record. It was not disproved. It was shown to be unmeasurable as stated, and it is withdrawn on that basis.
Bitcoin (COINBASE:BTCUSD), daily bars, measurement window 19 January 2015 to 20 August 2026.
"it seems when we poke above the 200 and the 200 is above 50 and 20, we drop soon after by a lot."
The observation was made on the daily chart. No other timeframe was checked by its author. The theory was stated on 20 August 2026 and the test specification was frozen (hash 0e35a54bde84) before any data was touched. The test concluded on 21 August 2026.
A daily high above the 200-day exponential moving average (EMA200), while the EMA200 sits above both the EMA50 and the EMA20 (a bearish stack), and the previous day's high was below the EMA200. Pokes within ten trading days of each other collapse into one episode, scored from the first.
Two choices were fixed before the data was seen:
Run on an hourly record aggregated to daily, the test found 561 usable days, 3 raw pokes and one episode after clustering. That is not a weak test. It is no test. Lower timeframes were then tried (4-hour: 19 episodes; 1-hour: 25). These were not part of the frozen specification and are reported for completeness only. Against the in-regime control, three of six measures carried the wrong sign.
The daily record is a COINBASE:BTCUSD export covering 4,232 contiguous sessions from 19 January 2015 to 20 August 2026, with no gaps, no duplicates and no open-high-low-close violations. It was cross-checked against an independent hourly record from a second exchange across 761 overlapping days: median close disagreement 0.036%, 90th percentile 0.112%, maximum 0.238%. The 561 usable days in attempt one are those 761 less roughly 200 days of warm-up for the 200-day average. The second exchange is not named and no series derived from its data is published.
This produced twelve episodes, eleven of them resolved.
Every measure below was run on a population that, as the finding further down shows, is mostly not the structure the claim concerns. All of it is withdrawn. The tables are printed so the work can be audited, not because they establish anything.
| Median maximum drawdown | n | 5 days | 10 days |
|---|---|---|---|
| Every day (after warm-up) | 4,027 | −3.60% | −5.10% |
| Every bearish-stack day | 1,214 | −3.52% | −4.92% |
| Bearish-stack, matched rally | 293 | −3.50% | −5.12% |
| The poke episodes | 11 | −2.61% | −2.80% |
In these eleven cases, poke days fell less than ordinary days, and the gap widens at ten sessions. It describes eleven cases and says nothing about the market.
A maximum drawdown is negative if price trades below the close even once in the window, so a negative reading carries almost no information on its own. That every episode shows one is therefore not a finding.
| Five sessions later | Closed higher | Median furthest rise |
|---|---|---|
| Random day | 54.7% | +3.96% |
| Poke day | 81.8% (9 of 11) | +5.87% |
Nine of eleven closed higher. This is a count of eleven cases drawn from a population that is mostly the wrong structure. It is not read as a signal, and no significance is claimed for it.
| Five-day drawdown reaches | Every day | Bearish-stack | Pokes |
|---|---|---|---|
| −2.00% | 67.1% | 68.1% | 81.8% |
| −3.00% | 55.6% | 55.6% | 27.3% |
| −5.00% | 37.8% | 35.7% | 18.2% |
A small wobble became more likely than usual, and a serious fall became about half as likely. On eleven cases none of this is distinguishable from chance.
At twenty sessions, eight episodes rose and three fell.
| Measure | The 8 that rose | The 3 that fell |
|---|---|---|
| Days below the 200 | median 50 (6–282) | median 86 (69–183) |
| Maximum % below the 200 | median −29% (−55 / −4) | median −32% (−38 / −30) |
| Rally into the poke | median +38% (5–133) | median +38% (37–47) |
| EMA200 slope, 20 days | median −2.4% | median −1.9% |
The rally into the poke is identical to a tenth of a percentage point. Depth, duration and slope overlap.
Asked whether the episodes were impulse waves, the episodes were presented blind: numbered, shuffled on a fixed seed, dates and outcomes sealed, on a log price scale, with each window anchored on the low that began the advance so no panel leaked its own outcome.
Two of the twelve were impulses.
The event definition does not select for the structure the claim is about. The measures above ran on the eleven resolved episodes, of which at most two were impulses, so every figure in them describes a population that is overwhelmingly not the thing the claim concerns. Every number in Measures 1 to 5, including those that pointed in the author's favour, is therefore withdrawn.
Blinding is what caught it. Shown the dates, an experienced labeller knows what happened in July 2018, and the labels follow the outcomes.
Offered as description and carrying no weight at n = 2: both rose, by a median of +10.6% at twenty sessions, with a median dip of −2.24%. The nine non-impulses were up six and down three.
The three episodes that dipped hardest are exactly the three that ended lower. No episode fell that far and recovered. The dividing line sits between −7.81% (recovered, +11.26% at twenty sessions) and −8.89% (did not, −6.89%).
The run log does not record the horizon these two drawdown figures were measured over. It is longer than five sessions: Measure 4 shows only two of eleven episodes reaching −5% within five. A clean threshold with eleven points either side of it is what coincidence looks like. It is flagged, not claimed, and it has not been tested.
Withdrawn. Untested, not disproved.
The claim was never shown false. It was shown to be unmeasurable with an event definition that catches mostly the wrong structures. Moving from 561 days to 4,232 took the question from impossible to unresolved, and no further.
Invert the search. Build the population structure-first: find the impulses across the record, then ask what they do at the 200-day average, rather than finding pokes and hoping they are impulses. The bottleneck becomes impulse identification, which is judgement rather than script, so the shape is: generate candidate windows mechanically, label them blind in batches, then measure.
Fixed in advance of any such work: eleven cases split two ways leaves about five and six. Whatever falls out is a hypothesis to freeze and test on data not yet seen. It is not a result.
A theory survived its test. The analysis did not. The trader's observation was never shown to be wrong. What was shown is that the event definition built to test it catches mostly the wrong structures, so the numbers could not speak to the claim either way.
Blinding earned its place the first time it was used. The episodes were numbered, shuffled and sealed before anyone was asked what they were. What that caught was a fault in the house analysis, not in the author's theory. Shown the dates, an experienced labeller knows what happened in July 2018.
The record that survives is worth more than the result that did not. A validated daily series, cross-checked against an independent exchange, and a blind-labelling protocol that works.
The rule we are keeping: state the population before measuring it, and check that it satisfies the claim's own preconditions.
Related reading: The Alderlux Null Test, When Bitcoin Moves Most and Bitcoin and the US Data Calendar.
Alderlux provides education and information only. We do not provide regulated financial advice, investment recommendations, legal or tax advice. Alderlux is not authorised by the Financial Conduct Authority. Past performance is not indicative of future results.