How we score a market setup 0–100

Smart Setups ranks what is happening across a few thousand markets and gives each one a number between 0 and 100. This is what that number is made of: every component, every weight, the two multipliers that can cut it, and the shorter rule that ranks a setup once it has fired.

What a setup is, and what we refuse to publish

A setup is a named structure the detectors found on one market and one timeframe: a breakout, a retest, a bounce off a level, a failed break, a trend continuation. The board scores structure, not outcome. It is telling you what the chart is doing, not what it will do next.

It is worth being blunt about the limit before the mechanics. The worker computes an implied trade plan (entry, invalidation, target), and the Risk component scores its geometry. None of it is published. Three fields are stripped server-side for every audience, free or paid, and the win rate is stripped too: the historical track record exists but every one of its routes is admin-only.

The event decides what the trade is

The scorer starts from events rather than from indicators. There are thirteen kinds, each with a base strength: a pattern breakout opens at 100, a breakout retest at 90, a trendline break at 85, a level or trendline bounce at 80, a reversal divergence at 65, a failed breakout at 55, a squeeze at 50, a level test at 40. The strongest one decides what the setup is called; each of up to two others leaning the same way adds ten points.

Two adjustments matter more than the base numbers. Strength halves every five bars and an event older than twelve bars is discarded outright, so the board cannot keep selling you something that happened last Tuesday. And a breakout on dead volume is damped toward half strength, while one that has held for two closed candles is lifted by ten percent. A break nobody participated in is not the same event as a break that stuck.

The arithmetic: five components, two gates

The five components are combined as a weighted blend, then two multipliers, then a clamp to 0–100.

• Context: 0.27. How much of the chart agrees with the direction: trend, momentum, volume, structure, volatility.
• Event: 0.27. The strongest thing that actually just happened, plus a bonus for up to two others leaning the same way.
• Participation: 0.18. Relative volume, taker flow and liquidity tier, in a 45/30/25 mix.
• Risk: 0.12. The geometry of the implied plan. The smallest weight, on purpose.
• Alignment: 0.16. Whether the timeframes above this one agree.

The two multipliers are the interesting part. A trend-continuation setup sitting in a range is cut to 75% of its score, but a breakout or a bounce is not, because compression is the condition those setups need rather than a defect. And any market whose last closed candle is more than two timeframes old is halved, which is what keeps a halted or delisted book from sitting at the top of the board on stale prices.

Context is capped rather than summed freely. Evidence is grouped into five categories and no category can contribute more than six units of weight, so a single fact restated six ways cannot carry a score. Full evidence credit is reached at seven weighted signals; below that the component is scaled down, so a setup with one lonely piece of confirmation reads as what it is.

When a component genuinely cannot be measured, its weight is dropped and the others are renormalised; it is not scored at a neutral 50. That has one consequence worth knowing: a weekly setup has no higher timeframe to align against, so the Alignment weight is removed and a 1w score is built from four components rather than five. It is on the same 0–100 range as a 1h score, but it is not quite the same measurement.

Fired setups are ranked on what has worked

Since 15 October 2026, a setup that has fired (a breakout, a trendline break or a retest) is no longer ranked by that blend. We record every one of them and measure how it played out: whether price moved 2% its way before it moved 2% against it, within 24 candles. Set against the other setups that fired at the same moment, so that a rising or falling market is not mistaken for skill, the blend did not separate the setups that worked from the ones that did not. Mostly it ranked setups by how well they agreed with the current trend, which paid in some weeks and cost in others.

Two things did separate them, both in the weeks before our October detector changes and in the weeks after: the type of setup, where a retest beat a fresh break by four to five points of race wins, and participation. Thin participation hurt both kinds; above that a retest did about as well however busy it got, while a fresh break did worse again when participation ran hot, which reads as a blow-off. So a fired setup’s score is now just those two.

• A retest with healthy participation scores about 81. Healthy means 35 or more on the 0–100 participation scale. Inside that, a fraction of a point comes off the further participation sits from 50, only so that equal setups do not tie.
• Thin participation costs 0.8 points for each point below 35. So a retest at 20 scores about 69.
• A fresh break scores 25 points below a retest. The same for a pattern breakout and a trendline break.
• A fresh break running hot loses 0.4 points for each point above 50. A retest is not marked down for heavy participation.
• The stale-data multiplier still applies. A market whose last closed candle is more than two timeframes old is still halved.

We checked it on days held back from the fit: among setups going the same way at the same moment, the higher-scored half won that race about four points more often than the lower half, where the old blend separated them by less than one. The five components are still computed and shown on every setup; on a fired setup they describe it, they do not rank it. Setups that have not fired yet were recorded the same way, and since 27 October they are ranked on that record too.

Setups that have not fired are ranked the same way

From 15 October we also recorded the setups that have not fired: breakout watches, support bounces and resistance fades, reversal attempts and trend continuations. Each one is entered at the last closed price while it is on the board, at most once every 24 candles per market, and runs the same race.

We read that race differently this time. A race that never moves 2% either way counts as a loss, and that rewards a market for being busy rather than for being right. So we counted only the races that moved, and asked how often price went the setup’s way first, against the other setups’ markets at the same moment. Among the common types, retests came out best, about 1.4 points above that line. Breakout watches matched it, support bounces and resistance fades were about a point below, reversal attempts about four below and fresh breakouts six to eight below. The order was the same on the 15-minute, hourly and four-hour boards, and in both halves of the period.

Inside a type, the blend still separated breakout watches: the best-scored fifth beat the worst by three to five points, mostly through context and higher-timeframe alignment. It did not separate the other types, and for bounces it pointed the wrong way in the first half of the period. So each type now has its own level, on the same scale as the fired setups.

• A breakout watch scores about 71 at a middling blend. It moves 0.35 points for each point of the blend, from about 64 up to the cap.
• A support bounce or resistance fade scores about 68. The blend only breaks ties, by under a point either way. The same holds for the next two.
• A reversal attempt scores about 58. That is below a bounce and just above a fresh break at its best.
• A trend continuation scores about 71 for now. Its own result was the best of any type, but on about 90 setups a day and with two losing days in eleven, so it ranks with the watches until a second period confirms it.
• Nothing scores above 81.5. That is the top of a retest with healthy participation.
• The stale-data multiplier still applies. The range cut for trend continuations does not: nothing in the record supported it.

On days held back from the fit, among all setups going the same way at the same moment, the higher-scored half won the race 4.2 points more often than the lower half, against 3.1 with the scores before. Fitted on the later days alone, the rule ranked the earlier days as well as before. The visible change is at the top of the board: bounces had scored 80 to 85 and some watches above 90, ahead of retests at 81, and they now sit below them.

Five minutes, and only closed candles

Nothing here is streaming. The scheduler runs a fixed five-minute loop, and every detector and indicator reads only candles whose close time has already passed. The public API then serves each query from a 45-second cache.

So the honest worst case, on a 15m board: up to fifteen minutes for the bar to close, up to five more for the next cycle, then the cache and the page poll. A 1d row is scored on a bar that can be twenty-four hours old by construction, and that is not staleness; it is what a daily candle is. If you need tick-level reaction, an alert is the right tool and this board is not.

The coverage number is smaller than “every market” too, deliberately. Three filters run before a symbol can be scored: fewer than sixty closed candles, a stablecoin or fiat base asset, and (for USD-quoted pairs) under fifty thousand dollars of 24-hour quote volume. On the box these figures were read from, Binance’s candle catalogue held 1,362 symbols and the 15m setup index held 1,220. About a tenth of the venue is filtered out before scoring begins.

What we are not claiming

• The score is a ranking, not a grade. Percentiles are computed per venue and per timeframe, within one five-minute pass. An 87 on Binance 4h and an 87 on another venue’s 15m are not the same claim.
• The stable-base filter has a hole we have seen fire. The base asset is derived by stripping known quote suffixes, so a fiat-quoted peg pair can slip through with a base the stablecoin list does not recognise. On one pass a tether/tenge pair was the highest-scoring 4h setup on the board. The mechanism is real and the fix is ours to make.
• Not every advertised venue is equally populated. The scorer accepts six venues; the boards offer five. On the stack these numbers came from, two of them returned no setups at all.
• Almost every threshold here is environment-overridable. The weights, the caps, the event half-life and the volume floor can all be changed on a running instance without a rebuild. The figures above are the code defaults, which is what this site runs.

None of that makes the number useless. It makes it a ranking of structural evidence, computed the same way for every market, that tells you where to look first. That is all it has ever claimed to be, and it is the reason the entry price is not in the payload.