What the indicator study measures

The indicator study measures one thing: the mean forward log return over the next 21 sessions when a signal is on at the close, compared to the mean when it is off, pooled across 18 miners. It also carries a caveat about why the p-value it prints is not a clean read on the evidence.

Start with one close. The fixture shows CleanSpark at its last session. The short exponential average is 9.961 and the long exponential average is 11.167. The MACD line is the short average less the long one:

9.961 - 11.167 = -1.206

That is one company, one timestamp. The study does not stop at one close. It gathers every close across every miner in the pool and sorts each into a group: the signal is on, or it is off.

For the low-RSI signal at the 21-session horizon, the two groups have these sizes:

• On: 1,060 sessions
• Off: 23,201 sessions

The mean forward log return in each group:

• Mean when on: 0.025
• Mean when off: -0.024

The gap between them, on its own line:

0.025 - -0.024 = 0.049

That gap is the quantity the test examines. The test is Welch's t, which does not assume the two groups share a variance. The t-statistic:

$t =$ 3.64

The p-value that follows from that t:

$p =$ 0.0003

Now the caveat, and it is the part that looks harder than it is. The windows overlap. A forward window of 21 sessions that starts at one close shares almost all of its sessions with the window that starts at the next close. Two consecutive windows of that length share every session except the first and the last. The 1,060 on-sessions are therefore not independent draws. They are correlated, because each window slides forward by a single session and keeps most of its content. The t-test treats the on-sessions as if they were independent. That can make the standard error too small, the t too large, and the p too small. The p as printed can overstate the evidence against the null.

The study reports the pooled p exactly as the test produces it. It does not apply a correction for the overlap. The figure is what the test gives under its own assumptions.

So: the study measures the gap between two means, one for signal-on closes and one for signal-off closes, pooled across the miner pool. The gap is 0.049. The p-value attached to that gap is computed on windows that overlap, so it sits lower than a p computed on truly independent draws would. The number describes the test as run, under the assumption of independent observations.