Support and resistance that holds: five rules, one test, and the cases where it fails

A support or resistance level is a price at which something happened more than once: a low that held, a high that capped, a round number that trades were executed around, a settlement price that positions were marked against. The line drawn on the chart is the analyst’s object. What the market did at that price is the evidence, and the two are regularly confused.

The confusion costs money in one specific way. A level drawn after the fact always explains the past, because the eye finds the touches that fit and the band widens until they do. A level whose definition was fixed before the count, and whose count is reported with it, can be checked by a second person and can be retired when it stops describing anything. That is the difference between a chart annotation and a measurement, and it is the whole content of this piece.

What follows is a test, not a promise. Five rules are fixed in advance — a band, a separation, a departure, a price basis, an as-of ATR — and then applied to nine candidate approaches to one level in one illustrative series. Four survive. The same nine approaches are then re-run at three band widths, which shows that the count is a function of the band and that the band therefore has to be published. The piece closes with the cases where the test fails and the sentence that keeps it honest: a level that was respected is a description of the past, and nothing in the arithmetic says it will be respected again.

What a level is, and what a line is not

Levels get drawn from a small set of sources, and each has a different claim on interpretation. A prior swing high or low is a price at which the market turned; a round number is a price at which orders cluster because people use round numbers; a volume-weighted price or a session’s VWAP is a price at which a quantity of business was done; a settlement or expiry price is a reference against which positions are marked; a prior consolidation’s edge is a price that separated two regimes.

None of those is a fact about the future. Each is a fact about the past with a mechanism attached, and the mechanisms differ. A round number has a reason to matter that a swing high does not — orders cluster there — and a swing high has a reason that a round number does not: it is where supply met demand and won. When two of these coincide, the level is said to be stronger, which usually means the analyst has found two reasons to draw the same line rather than one.

The line itself is where the trouble starts. Prices are continuous in principle and traders are not: a level is almost never touched at the same price twice, and the tolerance band — how close to the line counts as a touch — is doing most of the work of the count. A band of 0.25 ATR and a band of 1.0 ATR applied to the same series produce different levels, different counts and different conclusions, and both are defensible as drawings. Only one of them can be tested, and the test requires that the band be chosen before the touches are counted.

The test: five rules fixed in advance

The rules below are the ones this piece applies. They are not the only possible set; they are a set that can be written down, checked by a second person, and reproduced on another series. Each rule earns its place by catching a specific way the count inflates.

Rule 1 — the band, in ATR, not in price. A touch is a close inside the level ± half a band width, with the width set as a multiple of the ATR as of the touch date. ATR is used because a fixed number of ticks means something different in a quiet instrument and a volatile one, and because it lets the rule travel between instruments. At a level of 48.00 and an ATR of 1.20, a band of 0.5 ATR is 0.60 either side, so the band runs from 47.40 to 48.60 — 1.20 wide, 1.25 per cent of the level.

Rule 2 — separation. A new touch requires at least five bars since the previous counted touch. Without it, one episode of trading around a level becomes five touches, and five touches look like five confirmations. This is the rule that most inflates counts when it is absent, because the market rarely visits a level once.

Rule 3 — departure. A counted touch must be followed, within five bars, by a move of at least 1.0 ATR away from the level. A price that drifts into the band, sits in it and drifts out has not been repelled or rejected; it has passed through a zone slowly, which is the opposite observation dressed as the same one. The departure is measured from the level, not from the band edge, so that Rule 1’s width does not contaminate it.

Rule 4 — closes, not wicks. The touch is measured on the closing price. A bar whose high pokes into the band while the close finishes 0.20 below it has produced a wick, and a wick is a fact about one moment of the session rather than about where the session settled.

Rule 5 — the ATR as of the touch. The band uses the ATR computed at the date of the touch, never the ATR of today. A level re-tested in a calmer market has a narrow band; a level re-tested in a wider one has a wider band, and using today’s number for a touch two years ago retrofits the present into the past.

Table 1. The five rules, the inflation each prevents, the arithmetic that applies it, and what the note should report. Source: constructed for this piece.

The five rules are a checklist, and like most checklists their value is in the rejections. A count that reports only the touches that survived is not a test; it is a selection with a test’s vocabulary.

One level, tested

Table 2 applies the five rules to nine candidate approaches to a level at 48.00 in an illustrative series of 210 daily sessions — about ten months — with an ATR of 1.20 and a band of 47.40 to 48.60.

The candidates are the bars whose range came within the band; the table records each one’s date, low, high and close, its departure in ATR, whether it survives, and why. Every figure is computed, and the rejections are the interesting column.

Table 2. Nine candidate approaches to the illustrative level at 48.00 under five rules fixed in advance. ATR 1.20, band 47.40–48.60, separation 5 bars, departure ≥ 1.0 ATR. Source: constructed for this piece; illustrative, and no instrument or market is described.

The count is four. The arithmetic behind it is short enough to check by hand.

Eight of the nine closes fall inside the band; the ninth, day 74, closes at 47.20, twenty cents below the band’s floor, so it is a wick and Rule 4 rejects it. Of the eight, two are rejected by Rule 2 — days 14 and 41, each three bars after a counted touch, which is the same episode continuing rather than a fresh test. Of the six that remain, two are rejected by Rule 3: day 121 departed 0.70 ATR and day 156 departed 0.90, both under the one-ATR floor, which is the signature of a price grinding through the zone rather than leaving it. Four survive: days 11, 38, 96 and 191.

The departures of the four are 1.55, 2.05, 1.70 and 2.65 ATRs. They sum to 7.95, so the mean is 1.9875 ATR and the median is 1.875; in price, at an ATR of 1.20, they are 1.86, 2.46, 2.04 and 3.18, with a mean of 2.385. The smallest departure in the set — 1.86 — is more than three times the band’s half-width of 0.60, which is the property that makes the counted touches distinguishable from noise around the line: the price left the band by a distance larger than the band itself in every counted case.

The spacing is the other output, and it is the one that decides whether the level is worth the attention. The four counted touches fall on days 11, 38, 96 and 191: gaps of 27, 58 and 95 bars, one counted episode every 52.5 bars across the 210-session window, which is about one every two and a half months. A level tested four times in ten months is a level that the market has visited, not a level that defines the ten months. The note should say which, because “a well-tested level” and “four touches in ten months” are the same fact reported at two different levels of honesty.

The count is a function of the band

The single most useful thing to know about a level count is how easily it moves. Table 3 runs the same nine approaches through the same rules at three band widths, keeping Rules 2 to 5 unchanged.

Table 3. The same nine candidate approaches and the same rules at three band widths. Only Rule 1’s multiple changes; the separation, departure, close-basis and as-of rules are unaltered. Source: computed from Table 2.

The count doubles from two to four as the band doubles from a quarter to half an ATR, and rises again to five at a full ATR, where the wick on day 74 becomes a candidate with a close inside the wider band. Nothing about the market changed between the three rows. The band is a choice, the count is its consequence, and an analyst who tries several widths and reports the one that tells the better story has not measured a level — they have searched one.

The remedy is procedural rather than statistical: choose the band before the count, publish it with the count, and report the count at the neighbouring widths so the reader can see the sensitivity. A level that survives at 0.25 ATR and at 1.0 ATR is a different kind of object from one that appears only in the middle, and the counted column of Table 3 is what makes the difference visible.

Where the test fails

The test above is a filter, and it fails in five recognisable ways. Each is worth naming in the note, because each produces a count that looks like a measurement.

The band chosen after the touches. The commonest failure and the easiest to commit without noticing. Draw the line, look at the price, widen the band until three touches appear, then report three touches. Table 3 is the arithmetic of why this is a search rather than a measurement: a wider band admits more candidates by construction.

The level redrawn after a break. A price passes decisively through the level, and the line moves to the next swing so that the failed level is no longer on the chart. Every level survives if the ones that failed are removed, and the count of a level that has been redrawn three times measures the redrawing.

One session counted as many touches. Thirty prints at one price in one session is one episode. So is an hour of trading inside the band. Rule 2 handles this at the bar level, and it has to be applied at the session level too: a level “tested five times this week” is often one episode that lasted five days.

The departure that is the level working, or the market leaving. A large departure is genuinely ambiguous. It can mean the level was respected with force, or that unrelated news arrived twenty minutes later and the price went somewhere else. The test cannot distinguish the two, because it does not observe reasons; it observes prices. That ambiguity belongs in the note rather than in the count.

The regime that makes the level irrelevant. A level formed when the instrument traded at a different volatility, in a different session structure, or under a different set of participants is a line drawn on a market that no longer exists. Rule 5 keeps the ATR honest about the touch date, and it does not address the deeper problem: the level’s mechanism — the orders that clustered at that price — may have gone with the regime.

What the count does not tell you

The four counted touches are a description. They are not a probability, and the temptation to convert them into one is the reason a test like this needs a limitation section rather than a conclusion.

Four observations cannot support a rate. A count of four gives no usable estimate of how often the level is respected, because the denominator — the occasions on which the level might have been tested and was not approached — is unknown, and because four is not a sample from which a distribution can be inferred. Reporting “three of four held” would be a rate computed on four events chosen by a rule that was itself chosen after looking at the series.

The test also measures the wrong object if the question is about trading. It says whether a level is a well-defined, repeatedly-visited price with distinct episodes and non-trivial departures. It does not say whether a position taken at that level would have been profitable, what the stop would have cost, or how many times the level would have been crossed while a position was held. Those are questions about a strategy, and answering them requires a specification, a cost model, an execution rule and an error bar, not a count of touches.

And the sample is small by construction. A rule strict enough to be meaningful produces few counted touches; a rule loose enough to produce many produces touches that are not distinct. That tension cannot be optimised away, which is why the honest output of this test is a verdict about usability — is this level defined tightly enough that two analysts would draw the same line and count the same episodes — rather than a statement about what the price will do at it.

What a level test costs

Three engine calls produce the table: chart, for the swing structure the candidate list is drawn from, indicator, for the ATR as of each touch date, and verify to cross-check the price series against a second observation — a series adjusted differently is a different set of closes, and the closes decide the count. At the published $0.006125 on Base that is $0.018375, published rather than measured while the analyst services are pending review.

Two model calls draft the note at DeepSeek’s measured $0.00700: one to state the rules and the band before the count, one to write the table and the paragraph that reports what the count does not establish. That is $0.014000, so one level test is $0.032375.

Eight a month, for a book of instruments under review, is 24 engine calls ($0.147000) and 16 model calls ($0.112000): $0.259000, or $0.032375 a level and $3.11 over twelve months. At the metered ceiling of $0.00955 the model line rises to $0.152800 and the month to $0.299800, 15.8 per cent more; on Qwen at a flat $0.00875 the model line is $0.140000 and the month $0.287000, 10.8 per cent more. A level is five paid calls, so the 0.50 USDC opening balance — seventy-one calls at $0.00700 — holds fourteen level tests. The document parts that render the chart carry no published per-call price and sit outside every total here.

A decision rule

A level enters a note with five things attached, and the fifth is the one that keeps it from becoming a prediction.

Publish the level and the band together, in ATR. A level without a band is a drawing; a band without an as-of ATR cannot be reproduced. State the ATR, the multiple and the resulting prices.

Publish the count with its rule set and its rejections. Every candidate with its date and its verdict, as Table 2 does. A count of four with two separation rejections and two departure rejections is checkable; a count of four is not.

Publish the spacing. The gaps between counted touches, in bars, are the measure of whether the level describes the window or decorates it. One episode every two and a half months and one episode every three weeks are different claims about the same line.

Publish the sensitivity. The count at the neighbouring band widths, as in Table 3, so that a reader can see whether the level exists at one width only.

Then publish the invalidation. Not a prediction — a condition. For a position taken at the level, the distance from the entry to the far edge of the band is the distance at which the level’s description has stopped being true, which on this example is 0.60, or half an ATR, below 48.00. That number belongs in the note before the position exists, because it is the one figure the reader will need when the price is moving and the analyst is not thinking clearly.

What the note does not contain is a statement that the level will hold. The test produces a description of what the price did at a price, under rules fixed in advance, with the rejections shown. Whether that description is worth acting on is a question about the rest of the strategy — its entry, its size, its costs, its error bar — and it is answered somewhere other than in a chart.

Method note

The instrument, the level, the nine candidate bars, the 210-session window and every figure in Tables 2 and 3 are constructed for this piece and describe no market and no instrument; they exist so that the rules can be checked by hand. The band is 0.5 × 1.20 = 0.60 either side of 48.00, so 47.40 to 48.60, 1.20 wide and 1.25 per cent of the level. Rule 2 requires five bars since the previous counted touch, which rejects days 14 and 41 at three bars each. Rule 3 requires a departure of at least 1.20 in price within five bars, which rejects days 121 and 156 at 0.70 and 0.90 ATR. Rule 4 rejects day 74, whose close of 47.20 is outside the band. The four counted departures are 1.55, 2.05, 1.70 and 2.65 ATR — 7.95 in total, a mean of 1.9875 and a median of 1.875 — and in price 1.86, 2.46, 2.04 and 3.18. The spacing is 27, 58 and 95 bars across 210 sessions, one counted episode every 52.5 bars. Table 3 recomputes the same nine candidates at band multiples of 0.25, 0.50 and 1.00 with Rules 2 to 5 unchanged, and the admitted, counted, separation and departure columns are sums over that recomputation. Engine prices are the published $0.006125 per basic operation on Base from catalog.json, read 31 October 2026, published rather than measured because the analyst services are pending review and return no 402 quote. Model prices are measured from the live 402 handshake’s exact amount on Base on 31 October 2026 at a short prompt and max_tokens: 1000: $0.00700 for deepseek-v4-flash, $0.00955 at about 3,000 tokens, and $0.00875 for qwen3.8-flash at both lengths. Monthly counts are planning assumptions. Nothing here is a forecast of price, a statement that any level will hold or fail, a claim about the probability of any outcome, or investment advice; the count is a description of one constructed series.

The pillar is Finance and markets on x402 — three engines, four jobs, one price. The role that produces a level note is trade_analyst, whose brief leads with the setup and the levels, then the evidence, then what would invalidate the view, and which will not produce a forward-looking return. The compute is bought at x402.press/services/deepseek/ with a second model at x402.press/services/qwen/; the measured handshake is reviewed at DeepSeek: per-call chat completions over x402, and the two endpoints against each other at Inference endpoints compared. A level in a chart and a level in a curve are the same object seen twice, and Reading a yield curve is the version of this discipline applied to a set of tenors. The habit of fixing the reference set before comparing anything is the argument in Comparable company analysis template, where the peer set is chosen by rule rather than by fit.

x402.review is written and edited by N.P. Vincent. The publication is connected to x402.press, which sells the services reviewed here; that interest is disclosed on every piece that touches a service, and it is stated here rather than left to inference.

Filed under Jobs · Tags: support-resistance, charts, levels, trading