Backtesting Arena

Backtesting Arena

Back to blog

Coin Days Destroyed as a Bitcoin Signal: The Pattern Is Real. The Edge Is Not.

Rising Bitcoin while long-term holders sit still counts as deeply bullish on-chain. We replicated the pattern — and asked the one question missing: compared to what?

Backtesting Arena·September 3, 2026·8 min read·1 views
Coin Days Destroyed as a Bitcoin Signal: The Pattern Is Real. The Edge Is Not.

When Bitcoin rises sharply while long-term holders leave their coins untouched, on-chain analysis reads it as deeply bullish: the market is absorbing supply without old hands having to provide it. In late August 2026 that setup was in play again — Bitcoin up 23.6 % in fourteen days, the long-term-holder share of destroyed coin days down from 92.2 % to 86.2 %. Cited as evidence: nine such setups since 2016, with Bitcoin substantially higher a year later after almost every one.

We recalculated it. The replication works — and that is exactly where the problem starts.

What the metric measures

A bitcoin sitting still for one day accrues one coin day. When it moves, every coin day it accumulated is destroyed in that instant. The sum of that destruction is called Coin Days Destroyed. It weights movement by age rather than size: a single ten-year-old coin destroys as many coin days as 3,650 coins that arrived yesterday.

Split that destruction by holding period and you get the share attributable to long-term holders. When that share falls while the price rises, the story practically tells itself: the upward pressure is not coming from old supply hitting the market.

Two caveats belong right here. First, coin age measures time since the last movement on the blockchain — not how long anyone has owned anything. A wallet migration or an exchange deposit destroys coin days without a single sale; an exchange-internal sale changes owners without touching the chain at all. Second, the line between "short" and "long" is a convention. Our data source draws it at 150 days, the better-known Glassnode definition at 155. The two cohorts are similar, not identical.

The replication works

We built the condition exactly as the post defines it in its own footnote: a fourteen-day price gain of at least 23.6 percent, together with a drop in the share of at least 3.6 percentage points, then a 180-day cooldown before the next hit can count.

The result lines up. Since 2016 our count finds nine setups. The median twelve-month forward return is +90.0 %, and seven of eight completed windows are positive. The ninth setup is the current case from 26 August 2026 — its window is still open and enters no statistic.

The anchor matches too: on 28 August we measure a fourteen-day gain of 23.3 % and a share of 86.7 % after a drop of 5.7 percentage points. The data behind the viral chart is real, the arithmetic is reproducible, the pattern exists.

Stop here and you have a bullish signal.

Compared to what?

The question the chart never asks is the one that matters most: what are we comparing those +90 % against?

The intuitive answer — zero — is wrong. Bitcoin has gone up almost regardless, since 2016, if you wait long enough. So we ran the same calculation for every single day since 1 January 2016, with no condition at all: buy, wait twelve months, record the result.

On 70.9 % of all days Bitcoin was higher a year later. The median across all 3,531 evaluable days is +67.6 %.

That is the base the signal stands on. "Seven of eight positive" sounds impressive until you know that seven out of ten arbitrary days are positive too. And +90 % against +67.6 % is a difference — but not a landslide, not across eight observations.

The test that takes the setup apart

A setup built from two conditions raises an obvious question: which of the two carries the result? You can measure that by deleting one of them.

So we dropped the on-chain leg and kept only the price momentum: a fourteen-day gain of at least 23.6 %, otherwise identical counting and identical cooldown. No statement about long-term holders whatsoever.

That control arm produces fourteen setups and a twelve-month median of +95.8 %, with ten of thirteen windows positive.

So it does not come in lower than the full setup. It comes in higher. Over twelve months, the plain observation "Bitcoin just rose sharply" explains the same numbers as the elaborate on-chain construction. The leg that makes the story interesting is not the one carrying it.

The strongest counter-argument

In fairness: on shorter horizons the full setup does look better. After thirty days its median is +23.5 % against +9.6 % for pure momentum; after 91 days, +31.7 % against +16.5 %. One could argue the on-chain leg does contribute — just on a shorter timeframe.

That is a serious objection, and it still does not hold. For one thing, it is eight observations against thirteen; a median built from eight values shifts as soon as a single case lands differently. For another — and this weighs more — the thirty-day horizon was never the claim. The post argues in twelve months. Measuring first and then picking the horizon where the result looks best reproduces exactly the problem it was meant to solve.

The same applies to the thresholds themselves. "At least 23.6 percent" and "at least 3.6 percentage points" are not round numbers somebody fixed in advance. They are the coordinates of the current day, declared into a condition after the fact. That today's case satisfies the condition is therefore not a discovery but a construction.

One case that inverts the story

Among the nine setups there is one that appears in no summary.

On 10 October 2021 Bitcoin had gained 27.2 % in fourteen days, and the long-term-holder share of destroyed coin days had fallen by 14.7 percentage points — the steepest drop of all nine cases. By the logic of the signal, that was the most convincing setup in the entire series.

Bitcoin's highest daily close of 2021 landed on 9 November. That is exactly thirty days later. Twelve months after the signal, Bitcoin stood 65.2 % lower.

This does not make the pattern worthless. It shows what a median of eight values conceals: the very setup quoted as bullish confirmation once stood immediately in front of a cycle high.

How many cases are there really

A final point concerns counting. The cooldown is 180 days, the measured horizon is 365. Two setups six months apart therefore share half of their future — their results are not two independent observations but one with an overlap.

Count only cases a full year apart and nine setups become six, five of them with a completed twelve-month window. The median is then +84.2 %, all five positive. That is still respectable — it is also five cases. Our house rule: under thirty observations, a result is an anecdote, not evidence. Five is far below that.

What to take away

The finding is not "the signal is wrong". It is this: the pattern exists, and its predictive power cannot be told apart from two much simpler things — Bitcoin's general upward drift, and the fact that the price just rose sharply.

That turns into three questions worth asking of any signal chart that shows up in your feed:

Compared to what? When a hit rate is quoted without a base rate beside it, half the information is missing. For Bitcoin on a twelve-month horizon that base rate is high — 70.9 % since 2016.

What happens if I remove the interesting leg? Almost every composite setup has one simple component and one exciting one. If the simple one alone does the same work, the exciting one is decoration.

How many independent cases are there? Not how many dots are on the chart, but how many do not overlap. And where do the thresholds come from — were they fixed in advance, or are they today's coordinates?

None of which makes the share itself worthless. It is context: it tells you how old the thing currently moving is. We published it as a chart with full history back to 2009, with the total amount destroyed alongside it — because 86 % of a quiet market means something different from 86 % of a frantic one. It also carries what we noticed while building it: across the full history this share drifts structurally upward from around 10 % to roughly 90 %, simply because an ageing blockchain holds ever more old coins. Compare it against price across decades and you are mostly measuring the age of the network.

This is not our first check of this kind. We have described four of our own studies that failed at the same hurdle, and we locked the rules in advance for the Wyckoff Spring precisely to avoid this error. The difference here: this time the claim was not ours.

The open question we actually care about: which on-chain signal survives its own control arm — the version where you delete the on-chain leg and keep only the price? If you know one, name it. We will run the numbers.

Frequently asked questions

What are Coin Days Destroyed? Every bitcoin accrues one coin day per day it sits still. When it moves, all accumulated coin days are destroyed. The metric therefore weights movement by the age of the coins that moved rather than by their quantity.

Does a falling LTH share mean long-term holders have stopped selling? Not necessarily. Coin age measures time since the last on-chain movement, not duration of ownership. Wallet migrations and exchange deposits destroy coin days without a sale; exchange-internal sales change owners without leaving a trace on the chain.

Why does a base rate matter so much? Because it supplies the yardstick. Since 2016 Bitcoin has been higher a year later on 70.9 % of all days, with a median of +67.6 %. Every hit rate claimed for a bullish signal has to beat that number, not zero.

What is a control arm? The same measurement with one condition removed. If the result does not get worse, the removed condition contributed nothing. Here, pure price momentum delivered a twelve-month median of +95.8 % against +90.0 % for the full setup.

Why are nine events too few? Because with a median of eight values a single different outcome moves the result noticeably — and because at 180 days apart the twelve-month windows overlap. At a full year of separation, six events remain.

Is the setup refuted, then? Refuted would be too strong. Established it certainly is not: its result is fully explained by Bitcoin's general upward drift plus price momentum, with the on-chain component adding nothing on top.


Not investment advice, not a recommendation, not a forecast. Historical patterns are no promise about the future. All figures as of 31 August 2026, data source Bitcoin Research Kit; the analysis is stored as a verification script in our repository.

Study the Past — Improve your Future 🥋

Try it yourself

Run the backtest with your own parameters and time ranges.

Run backtest →
📬

Don't miss new blog posts

One short email per new post — strategies, backtests, market analysis. No spam, unsubscribe with one click anytime.

By subscribing you accept our privacy policy. We use Resend for delivery. Double opt-in confirmation required.

Comments (0)

Join free to post comments.

Sign up →

No comments yet. Be the first!