How a sticky-threshold regime model beat a falling COIN by staying long and selective

Over a window in which COIN actually fell, a long-only rule still finished comfortably in the black by being selective about when it engaged and sticky about when it let go. That rare outcome is what makes this one worth reading. It is a backtest of one fixed rule set replayed through COIN's saved history, not a forecast or a recommendation, and the live sample is still too small to grade.

COIN's standing in its own percentile range is the regime this model reads, but no single trigger ever settles the call. An upper level turns a regime on and a distinct, lower level turns it off, so a fleeting dip back over the line leaves the position untouched, the deliberate hysteresis that calms whipsaw right at the edge. Every entry this window pointed the same way, twenty-four long trades and zero shorts, and two thirds of them paid, a 66.67% win rate that at one repeated stake came to +135.44%.

How the model is built, end to end.

The reading the model lives by is where COIN sits in its own percentile range, and the twist is that it carries memory. A single level decides when to enter a regime; a different, lower level decides when to exit it. Between those two levels the model holds its current view rather than reacting, so the small back-and-forth that would whipsaw a single-threshold rule simply passes underneath it. The regime only changes when the reading travels far enough to clear the second threshold, which keeps the model committed through ordinary noise.

The regime read grants permission; the sticky thresholds decide how long that permission lasts. Once a long regime is established the model opens with a stop and a target set in advance and stays with it until either level is hit or the reading finally crosses the lower exit threshold. On COIN that produced a fairly even spread of outcomes: the target, the stop, and the signal exit each closed a meaningful share of trades, with the settling regime read accounting for the largest single group of clean exits after the target.

The illustration above walks the model across COIN in a single line: the percentile reading, the two thresholds that frame the regime, and the long position they permit. The point of the design is restraint at the boundary, holding a view through small reversals instead of trading every one of them. The model's lineage belongs in the picture too. It began as one option among many in an automated search and reached deployment only by clearing backtest and walk-forward checks, the full route from idea to working rules behind every figure here.

At a glance

These figures come from a backtest of the model on COIN, scored against fixed acceptance gates, not from a live track record.

Walk-forward verification

A trade walked through

The walked example is a long held about two days. With COIN's percentile reading inside the long regime, the model bought near 254 dollars and the target filled near 303 dollars for +19.64%. It is the model working as intended, a regime held steady through the entry and an exit at the target once the move arrived.

What the full trade record shows

Across its 24 COIN trades the model won 16 and lost 8. The exits spread fairly evenly: 8 reached the profit target, 7 were stopped out, 6 closed when the regime read crossed the lower exit threshold, and 3 closed on a time exit.

A balanced split like this fits a sticky-threshold model: with whipsaw damped at the boundary, trades tend to resolve deliberately, some at the target, some at the stop, and a solid share when the regime read finally clears the lower exit level.

The biggest winner was a long that ran to its target for +20.27% over about five days in June 2025, opened inside a firmly established regime. The selective, long-only book meant most winners were patient holds carried until the target or the regime exit, not quick scalps.

The worst trade was a long stopped out for -6.44% over roughly eight days in April 2026; that single loss is also the equity curve's deepest drawdown, narrowly above the drawdown gate's limit and the reason that gate did not clear.

No single trade carries the record, and the standout of the book is what the sticky thresholds achieved as a whole, a gain on a stock that fell, earned by holding a selective long view through the noise rather than trading every reversal.

How does this compare to just holding COIN

Over the same window the model was tested on, simply buying COIN and holding would have lost money, so here the comparison flatters the rule rather than the stock. Setting the two together shows the model did the thing a passive holder could not, it profited while the shares fell, and the tiles below put numbers on that separation.

How well does the model reproduce its tape?

Walk-forward verification checks whether the saved rule path reproduces the expected signal behavior on held-out timestamps it was not built on. It is a consistency and replay-integrity test, not proof of live profit. A clean reproduction means the deployed rules behave like the studied ones; it says nothing about whether COIN will keep offering the regimes the sticky thresholds were tuned to read.

In live trading the model has been quiet so far, with too few signals to set beside the backtest. Until more live trades accumulate, the backtest is the only evidence on hand, and it should be read as exactly that, a study of how the rules behaved on saved history.

When this approach fails

The model's failures come from the same mechanism that protects it, the sticky thresholds. The same memory that carries it through noise can also keep it long a beat too long when a regime is genuinely turning, which is how the deepest trade lost 6.44% on a long stopped out in April 2026. Being long only, it has no way to profit from a decline once a regime breaks, and a small sample of twenty-four trades means a single bad stretch weighs heavily on the record.

Three things are worth watching if this ever trades at size. The first is the modest sample, just twenty-four trades, which leaves the statistics thin and the sample-size gate unmet. The second is the long-only exposure, since the model cannot hedge or profit from a sustained fall once a regime breaks. The third is the gap between live and backtested behavior, the first sign the percentile regimes the thresholds were tuned to are no longer forming the same way.

Risk and honest limits

On this run the model's automated checks logged a caution rather than a clean pass. It is one more reason to read everything here as backtest evidence about the rules, not a verdict on the stock.

Lifecycle