Field Note 13 / Research Methods

The Coin Rose. What Happened Along the Way?

We studied 3.3 million hourly observations across 515 cryptocurrencies. The central lesson: finding a large rise and finding a movement that is convenient to trade are different research questions.

A coin can reach an impressive high, fall deeply below its starting price, and finish the week in the red. A screenshot of the high captures only one part of that story. Our Paths laboratory examined what happened around it.

More upside, more downside

The clearest example came from observations with high trading turnover and recent daily price growth. Over the following seven days, this group had a larger median maximum rise than the background reference. It also had a much deeper median fall below the starting price, and a greater share of negative weekly closes.

The movements became larger in both directions. The median maximum rise increased by 4.8 percentage points; the median maximum fall became 7.3 percentage points deeper. The share finishing below the starting price increased by 10.3 percentage points.

Each chart summarises a separate measurement across observations. These medians cannot be joined into the price path of a single coin. Maximum rise and maximum fall are both measured from the starting price; the latter is not a peak-to-trough drawdown.

What we actually measured

The laboratory used 3,332,403 hourly grid observations across 515 coins, without a signal filter. For each starting point, it measured the subsequent maximum rise, maximum fall, closing change, timing of the extremes, and time spent above the starting price. The underlying market history came from the Klines research dataset.

The full-population summaries already showed why the endpoint matters. Across complete seven-day windows, the median maximum rise was +7.0%, the median maximum fall was −7.5%, and the median closing change was −1.8%. The median share of time spent above the starting price was 36.1%.

These are separate summaries of a distribution. They do not mean that a typical executable trade earned 7%, then lost 7.5%, then closed at −1.8%. The order, timing and actual price path have to be measured for each observation.

The conditions were chosen before the later-period check

We examined 39 features available at the observation time. Their research-period deciles supplied the candidate boundaries. The research period ran from 1 November 2025 through 30 April 2026 and contained 1,842,674 observations.

Five candidates with the largest measured distribution differences were then frozen and checked on later data. The later period ran from 1 May through 16 September 2026, covering 1,474,262 observations over 138 days. The table above comes from that comparison, rather than the whole dataset.

This procedure separates candidate selection from the subsequent check. It does not make the later period a completely independent holdout: the project had previously examined those months in other work. We describe it as a stability check.

The distribution differences were small: a Kolmogorov–Smirnov distance of about 0.05 for the best single condition and 0.10 for the best pair. Those results support a limited observation about the measured paths, rather than a large change in the market's behaviour.

An earlier peak still leaves an exit problem

The high-turnover, recent-growth group reached its maximum before its minimum in 66.1% of observations, compared with 55.8% in the background. Its median time to the maximum was 32 hours. That makes the result more interesting than a simple claim that active coins only fall.

There was often a substantial rise along the way. But the historical maximum becomes known after the window has unfolded. Capturing it would require a decision rule available at the time, an executable entry and exit, and a cost model. This laboratory did not establish those.

A maximum favourable excursion is a price measurement. It is not an achieved return. A trader's result depends on when they entered, what happened before the exit, whether they could hold through it, and what prices were actually available.

Why the scale needs an asterisk

Neighbouring hourly observations share much of the same future week. They are not 3.3 million independent experiments, and they are not 3.3 million trades.

Activity was also concentrated. In the high-turnover, recent-growth group, the ten most represented coins accounted for 21% of observations and the ten busiest days for 23%. The reported median intervals therefore used a two-way coin-and-day cluster bootstrap, rather than resampling rows as if they were independent. The group’s intervals were +9.7% to +13.9% for maximum rise and −16.6% to −11.7% for maximum fall.

The historical window was a bearish market period, described in the report as 52% below its peak. That limits how far the result can be generalised. A different market period, universe or execution setting could produce different paths.

The only candidate that slightly improved the upside/downside balance involved low-turnover coins. Execution costs were not measured there. A milder historical path in an illiquid asset does not establish that a position can be entered or exited economically.

What this changes about the research question

It is tempting to group observations by the large rise they eventually produced and then describe what happened before it. That can be useful for understanding completed paths. It cannot, by itself, become a detector: membership in the group depends on a future outcome.

Our practical question becomes more demanding: which conditions available now change the whole subsequent path, including its adverse movement and endpoint? The Paths laboratory found an example where selecting more upside also selected more downside and more negative weekly closes. That is a useful result even though it supplies no ready-made trading strategy.

Looking beyond the attractive maximum helps keep the research honest. The movement worth detecting still has to survive the journey from an observed price path to a trade someone could actually execute.

Source: Karaptic Paths laboratory result, dated 25 September 2026. This article summarises the historical findings and does not reproduce Klines production formulas or trading settings. Background: What Is Left for a Retail Trader?, the Klines lab, and our volume lead–lag tests.