Net Unrealised Profit/Loss

The heaviest valuation signal this model ever carried, removed at a measured cost to its sharpest recent call — not because it scored badly, but because it is algebraically the same number as a signal we were already scoring beside it.

What it measures

Market capitalisation minus realised capitalisation, divided by market capitalisation. Realised cap values every coin at the price it last moved on-chain, so the result is the share of the network's total value that is currently paper profit. Above zero the average coin is held in profit; below zero the average coin is underwater. We scored it as a trailing percentile rather than against fixed levels, because the absolute bands compress every cycle.

• The bands we scored, from the bottom up: below -0.25 is deep capitulation, below 0 is the accumulation zone, 0 to 0.25 early uptrend, 0.25 to 0.5 mid-cycle, 0.5 to 0.65 late mid-cycle, 0.65 to 0.72 caution, 0.72 to 0.82 distribution, and above 0.82 euphoria.
• Those bands still compute and still print. What they no longer do is vote — the weight is zero, so the reading you see on this model is a lens, not an input.
• Here is the fact the rest of this page turns on. NUPL is not correlated with MVRV; it is DEFINED as one minus one over MVRV. NUPL 0.75 is MVRV 4.0. NUPL 0 is MVRV 1.0, the point where market price equals realised price. NUPL -0.25 is MVRV 0.8. There is no market condition in which the two disagree, because they are one quantity under two names.
• Realised price is the same variable a third time: realised price is price multiplied by one minus NUPL, so scoring it adds a third vote to the same reading.
• We were scoring all three at once — NUPL at 0.09, MVRV at 0.07 and realised price at 0.04 — and the contrarian-confluence block read the same variable a fourth time. One number, 0.20 of the model's weight, four touches.

Our data: Checkonchain, refreshed from the published NUPL chart and banked here as a frozen write-once daily series, non-revising, like every other graded input. Coverage runs from 2012-10-27. The data is not the limitation on this page and never was: the register's own signal-silence audit records this metric as zero-weighted by decision rather than absent, and names the file and the date range to prove the distinction.

What happened when we put it in the model

Is it independent of the valuation signal we kept?

No. It is the same condition, written differently. We measure a candidate's independence as the phi coefficient against the mvrv < 1.0 anchor, the valuation condition the model already reads. Net Unrealised Profit/Loss scores 0.9936. That is not a high correlation, it is an identity wearing two names, and we can show it on our own series rather than assert it from the algebra: the median absolute difference between NUPL and one minus one over MVRV is 0.0014 across 5,046 days. For scale, the same phi measurement, run against the equivalent form of this anchor, returns 0.465 for a miner-cost-basis reading and 0.412 for a long-term-holder conviction reading. Those are candidates from genuinely different on-chain axes. This is not one of them.

Recorded at brain/IMPROVEMENT_PROPOSALS.md:436.

What we concluded

Every other page in this family records a candidate that was tested, measured worse, and went to zero. This one is the opposite shape, and it is the more uncomfortable story. This signal was not a candidate. It was live, it was the single heaviest valuation input we had, and it had been passing for as long as we had been measuring.

It was passing because the measurement could not see the problem. A walk-forward test grades the composite the model produces. It cannot tell you that three of the inputs feeding that composite are the same number, because the composite it grades is the only thing it ever gets to look at. Collinearity is invisible to an out-of-sample score by construction — it does not make the model fail the test, it makes the model quietly place one fifth of its conviction on a single variable and call it three opinions.

What found it was not a better test. It was reading the definitions of two signals next to each other and noticing that one was the other rearranged. Nothing in the gate would ever have surfaced that, and the gate had been green throughout.

The correction was made in two steps and the second one cost us. Collapsing the triple count improved everything we measure and cost nothing — walk-forward geomean edge 1.046 to 1.090, current-cycle edge 1.55 to 1.99, and the 2025 top still graded A. Then removing the signal entirely improved the walk-forward again, 1.090 to 1.140, and took the 2025 top from an A to a C while the current cycle's dollar fold fell from 2.02 to 1.80.

We took the second step anyway, and the reasoning is the part worth arguing with. The A on the most recent top was the single most quotable thing this model had. Keeping it meant keeping a double-count that the older cycles said was hurting us — and a model that grades its most recent event best is describing the event it was most recently tuned near, which is the failure the walk-forward exists to catch rather than to certify. We traded the sharpest recent call for the more robust history, and we are publishing the price of that trade rather than the half of it that flatters us.

The honest generalisation, and the reason this page exists: passing a test is not the same as being true. A signal can clear every gate for years while the thing that makes it clear them is a defect the gate has no way to detect. If you are reading a track record anywhere — ours included — the question that separates a measured model from a fitted one is not how many tests it passed. It is what class of error those tests were capable of finding.

It stays wired and computed at weight zero, which is why both arms above can still be re-run from a single environment variable. We did not delete it; we priced it at nothing.

Where our own notes disagree

Other candidates we put through the same gate

Each of these was wired into the same model, measured against the same walk-forward gate, and written up with the numbers it produced — including the cells where it helped. A candidate can improve several measures and still be reverted, because the gate is a conjunction and buying one measure by selling another fails it.

Read the current call in today's briefing, or the timestamped record of every call in the call ledger.