Two versions of the same measurement are plotted, and on this instrument they are far apart. That gap is the whole idea.

UNCONDITIONAL, every regime, 3,001 overlapping windows:
Median 1,874
Core 50% 1,806.74 - 1,948.84 (7.6% of spot)
Wide 80% 1,724.25 - 2,033.74 (16.5%)

REGIME-MATCHED, only windows that began in conditions resembling today:
Median 1,875
Core 50% 1,823 - 1,936 (6.0%)
Wide 80% 1,768 - 1,992 (11.9%)

Filtering removes a fifth of the core width and 28% of the wide band. On Bitcoin this morning the same comparison barely moved the numbers. Here it moves them a lot, and the reason is printed two lines further down the panel.

THE BAND WAS DOCUMENTED AS TOO WIDE

Out-of-sample coverage, fitted on older history and tested on newer data it never saw: core 60.8% against a 50% target, wide 88.3% against 80%.

Both are over target. An interval that swallows six outcomes in ten while claiming five is not cautious, it is uninformative - and over-coverage is the failure mode that generates no complaint, because every result keeps landing inside and the scoreboard stays green. We found exactly this across our own resolved forecasts last week, published the diagnosis before the correction existed, and did not re-seal anything.

The unconditional band above is the version that produced those numbers. The narrower one is what happens when the sample is restricted to windows that opened in conditions like today's, measured at the open rather than at the close, because measuring at the close would leak the outcome being predicted.

WHAT IT COSTS

Sample size, stated rather than buried. The wide band rests on 125 genuinely independent windows; the filtered one on 41. Consecutive windows share 23 of their 24 bars, so quantile error bars scale with the smaller count, and 41 windows is thin evidence. A tighter band read from fewer observations is not automatically a better band, and anyone tightening a range should be asked what they gave up to do it.

DIRECTION: NONE WORTH TRADING

The regime line reads downtrend. ADX is 18.7, which is weak - well under the threshold where a trend reading means anything. That is a label with nothing behind it, and the median sits within 0.1% of spot.

The drawn path runs to the top of the filtered core near 1,937. It is one illustration of a route through the distribution, not a claim about direction, and it is the element of this chart most likely to be wrong. Our own 24-hour forecast for ETH, sealed and Bitcoin-timestamped this morning at 07:32 UTC on a shorter horizon than the four days plotted here, carries no directional lean at all: core 1,852 - 1,904.

THE SHAPE OF THIS DISTRIBUTION

Empirical width is 1.21x the textbook sigma-root-t band. The conventional formula is 21% too NARROW on ETH at this horizon - the opposite of what the same calculation does on Bitcoin, where it comes out too wide. There is no single correction factor that fixes both.

Skew is +0.10, which is as close to symmetric as anything we measure. Excess kurtosis 2.46, so the tails are fat even though they are even.

Widths scale almost exactly as diffusion predicts: 7.6% at 24 bars, 10.7% at 48 (1.41x, where root-t expects 1.41x) and 13.2% at 72 (1.74x against 1.73x). ETH behaves like independent increments at these horizons. Bitcoin, measured the same way this morning, widens materially faster. Same method, same day, two different animals.

Nothing here is a signal, a target or a recommendation. It is a description of how this market has moved before, with its probability stated, its sample size printed, and its known weakness named.

Educational content - not financial advice.