The Divergence Archetype: acting when two readings begin to move apart
Direction comes from the relationship between two monitored readings: the method sits flat while they move together and takes a position only at the moment they begin to move apart, with the primary reading setting the side and the second only needing to move the opposite way.
Method archetype
The Divergence archetype watches two monitored market readings at once and asks a single question of the pair: are they moving together, or are they beginning to move apart. While the two move in step it does nothing. It fires only at the moment they start to diverge, and the side of that signal is set by the reading it treats as primary.
The word that matters here is apart. This archetype is not built on the level of either reading, nor on a single reading reaching an extreme. It is built on the relationship between two readings, and specifically on the moment that relationship breaks step. That moment, not the value of either reading, is the whole signal, and it is why this article only ever calls the inputs a primary monitored reading and a second monitored reading.
Two readings, and what together means
Most of the method's life is spent doing nothing. As long as the two readings rise and fall together, there is no divergence to act on, so the method holds flat and waits. Moving together is the ordinary state of the pair, and the method treats it as a reason to stay out, not a reason to act.
This is worth dwelling on because it is the opposite of how a relationship is often used. A method that leans on two readings moving together would be busiest while they track each other and would step aside when they part. This method is the mirror of that: it is quiet while they track and only wakes up when they part. Being together is the off state.
The slopes: reducing each reading to the direction of its recent move
To judge whether the two readings are moving together or apart, the method reduces each one to the direction of its recent move, its slope over a recent window. A reading that has been rising has an upward slope; one that has been falling has a downward slope. The pair is moving together when their slopes share a sign and apart when their slopes oppose.
The asymmetry: the primary reading sets the side
The two readings do not play equal roles. One is treated as the primary reading, and its turn is what sets the direction: when the primary turns up while the second is falling, the method reads a long; when the primary turns down while the second is rising, it reads a short. The second reading never sets the side. Its only job is to be moving the opposite way, confirming that the two have genuinely split.
The onset: acting on the turn, not the duration
The method does not act for as long as a divergence lasts. It acts once, at the onset, on the single bar where the primary reading's slope crosses over to oppose the second. If the split then persists for many bars, the method does not keep taking the position again; the primary has already turned, so the fresh-turn condition is no longer met.
This makes the signal a stamp on a moment rather than a state it sits in. It marks the instant the pair breaks step, and then it is done with that event, which keeps the method from leaning on the same split over and over.
Three event readings, not held states
On each bar the method emits one of three readings: a long-event, a short-event, or flat. The difference from the stateful archetypes is the whole point of this section: a long or a short here is an event that marks the onset of a split, not a condition the signal keeps holding. It emits a long-event on the bar a primary up-turn opens against a falling second, a short-event on the mirror, and flat on every other bar, which is almost all of them.
This is why the same signal can be described as acting once and as reading flat almost always without contradiction: a long or a short is a one-bar mark at the onset, and the very next bar reads flat again even while the two readings stay apart. A divergence that lasts a long time still shows up as a single event at its start, so long stretches of flat are the method working as designed, not a sign that it has stalled. Whether a deployed model then holds a position after that event is a separate matter, covered further on.
No filter at the rule level: the relationship is the signal
Several of the other archetypes carry a quality and availability mask inside the construct that can stand the method aside. At the archetype-rule level this construct has no such separate filter: the reading comes straight from the two readings and their slopes. Some deployed models may still apply standard availability, quality, or risk controls around the construct, but the divergence rule itself adds none.
Not a correlation method: acting on the break, not the link
It is worth drawing the line between this family and a correlation or link method, because both watch two readings and a careless reader could blur them. A link method borrows direction while the two readings move together and steps aside when the link breaks. This method does the exact opposite: it is silent while they move together and acts only when they break apart.
The two are mirror images on the same picture. Where a link method is busy, this one is quiet, and the breakdown that sends a link method to the sidelines is the very event this method waits for. Sharing the surface, two readings, does not make them the same archetype; the trigger is inverted.
How a position ends
Because the signal is stamped on a moment rather than held as a state, what happens after the onset is a matter of how a deployed model chooses to manage the position, not of the construct re-reading the split. A deployed model may close a holding through standard protective exits, a protective stop, a profit target, and a maximum holding time. The exact levels and durations are model-specific and not part of this family-level description.
Validation and lifecycle: what a backtest does and does not show
Models built this way are judged through the same staged process as every other family: an idea, a backtest, and a held-back walk-forward before the method is relied on. A family-level explainer is not proof that every model in the family has cleared every gate.
The caveat: a divergence is a moment, not a forecast
There are a few caveats a reader must hold onto. A divergence is a description of two readings parting on a bar, not a forecast that anything in particular follows; many splits are noise and lead nowhere. Because the trigger fires on a fresh turn, a primary reading that wobbles back and forth can produce a run of turns and act repeatedly. And because the construct screens nothing, the signal is only as trustworthy as the two readings feeding it.
It also helps to be precise about direction. A long-event here is the method's own reading when the primary reading turns up against a falling second; it is a description of the method's stance, not a claim that the traded asset must rise. The honest way to read a method like this is therefore against a simple buy-and-hold benchmark, asking whether the positions a deployed model opens on these events added anything beyond what a passive position would have had. This explainer is number-less, so it makes no claim about whether any model clears that bar; it states the caveats so a divergence is never mistaken for a certainty.
How to read it on the Commentary Desk
The method's behavior maps onto three desk-level reads, all described in family-level terms.
Informational
Context, no action implied: the two readings are moving together, so there is no split and the method is in its routine flat state, or the readings are drifting but the primary has not turned. Note it as context, never as a live call.
Warning
An event worth flagging: the two readings have begun to move apart and the primary has turned on the bar, so a fresh long-event or short-event has just fired; or, at the deployed-model level, a protective close has fired on a position opened earlier.
Elevated caution
Flag, but interpret carefully rather than act mechanically: the primary reading is wobbling so the pair is splitting and rejoining repeatedly, or one of the two readings looks distorted, which the construct itself does not screen for.
What this archetype is NOT
- It is not a correlation or link method. It does not act while the two readings move together and step aside on a break; it does the opposite, staying flat while they move together and acting on the split.
- It is not a symmetric, two-equal-readings method. The primary reading sets the direction and the second only confirms the split by moving the other way; the two are not interchangeable.
- It is not a level or extreme method. There is no z-score, no percentile, no band; the trigger is the sign of two slopes opposing, not how high or low either reading sits.
- It is not a state the method sits in. It emits a one-bar event at the onset of a split and does not re-emit while the split continues; a long or a short here is a moment, not a held position.
- And the monitored readings are not necessarily price. The method works on whatever two market reads it is given, and this explainer keeps both masked throughout.
What we can and cannot claim
This explainer describes the family mechanism, not the wiring of any one model. Which two readings are watched, which one is treated as primary, and the length of the window the slopes are measured over are model internals, not stated here because they are chosen and re-checked per model rather than fixed constants the article is withholding. The monitored readings are not necessarily price, and this explainer never names them.
A method is not a recommendation, and a model built this way is not a guarantee of future results. It is a disciplined way of noticing when two readings begin to move apart and taking a position on that moment, and it earns its place only by surviving an idea, a backtest, a held-back test, and a staged review before it is relied on.
Sources
This article is based on Stonewell One research, including backtesting, walk-forward verification, deployment monitoring, and model-risk review.