Methodology
How this works
One indicator, the same math everywhere, and a plain account of what it can and can't tell you – including where it's wrong.
01 · Orientation
What this is
Arq watches trend direction, and trend changes, across three kinds of markets: crypto perps, stocks/metals/FX, and on-chain DEX tokens. It runs one indicator – the same math, the same rules – over all of them.
That's the whole product: not a signal service, not a predictor. A trend either flipped on a closed candle or it didn't. We try to say that plainly, show our work, and be upfront about the one thing this kind of indicator is structurally bad at (predicting whether a fresh flip will hold – see Fakeouts, honestly).
02 · The math
The indicator
Arq's primary signal is a Supertrend line built on a 10-period ATR with Wilder (RMA) smoothing, multiplier 3. That multiplier-3 lane is what every "Bull"/"Bear" state and every flip on this site refers to. A tighter multiplier-2 lane is also computed as secondary context in a few places (research notes, agreement checks) – it is never the source of a headline flip or an alert.
Mechanically: the indicator builds an upper and lower band around price, offset from the midpoint by 3×ATR(10). The bands only ever ratchet toward price, never away from it. Direction flips – bull to bear, or back – exactly when a closed candle's close crosses to the other side of the band currently in force. Nothing else moves the line: no news, no sentiment, no manual override.
No black box: this exact math is ported independently across the parts of the system that compute it, and a golden-fixture regression test pins a hand-built candle series' expected Supertrend values and flips so any future change that silently drifts the math fails the test immediately.
03 · Confirmation rule
Closed candles only
A flip only becomes real when the candle that crossed the line closes. Price can poke across the line mid-candle and pull back before the close – that's not a flip, and Arq never reports it as one.
While a candle is still forming, the UI shows the tentative cross as amber, labeled "unconfirmed," with a live countdown to the candle's close. It only turns into a real, colored flip if it's still crossed when that candle closes.
amber dot = forming candle, unconfirmed · line resolves once the candle closes
04 · Coverage
Timeframes & sources
Each market type gets candles from one source, at the timeframes that source can actually support well:
| Market | Source | Timeframes |
|---|---|---|
| Crypto (perps) | Hyperliquid | 1h4h1d1w |
| Stocks, metals, FX, indices | Yahoo Finance | 1d1w |
| On-chain DEX tokens | DexScreener (resolve) + GeckoTerminal (candles) | 1h4h |
Weekly candles are UTC Monday-to-Sunday weeks – aggregated from daily candles the same way for every source, so a "1W" flip means the same thing everywhere on the site.
On-chain candles are liquidity-gated before they're trusted: a pool without enough depth doesn't get a lane at all. And every new listing – on any source – needs 20 closed candles on a given timeframe before that lane reports anything. That floor is exactly twice the ATR(10) warm-up period: the indicator's earliest values are biased by how the smoothing window was seeded, so a lane younger than that would be showing you noise dressed up as a signal.
05 · The hard part
Fakeouts, honestly
Some flips reverse almost immediately. We call that a flip-back: a lane flips one way, then flips back the other way within a short window – the first flip turned out to be noise, not a real regime change.
We label a flip a flip-back retrospectively, once we can see whether it held. The window is measured in closed bars of that lane's own timeframe:
| Lane | Flip-back window |
|---|---|
| 1h | 6 bars (provisional – not yet separately calibrated) |
| 4h | 6 bars |
| 1d | 3 bars |
| 1w | 2 bars |
Here's the part that matters most: we ran a pre-registered study asking whether a fakeout could be predicted at the moment it flips – before you know how it turns out. Across 8,568 historical flips, we tested five candidate features (how deep the close crossed the band, whether higher timeframes agreed, whether the tighter multiplier-2 lane agreed, recent chop, and candle "quality"). We set the usefulness bar before looking at results – a feature had to split flips into groups whose flip-back rates differed by at least 2×, with at least 100 flips per group, and the split had to hold up on a held-out slice of history.
Nothing cleared that bar. The best-separating feature reached about 1.6×, and it worked in the wrong direction (flips that agreed with the higher timeframe flipped back slightly more, not less). So Arq does not publish a fakeout prediction, at flip time or otherwise – we don't think one would be honest to show.
What we show instead, both strictly after the fact:
- Amber flip-back markers on flips that have since reversed within their window.
- A per-asset "flip-quality" line on each asset page – flip count, flip-back rate, and median hold time for the selected timeframe – as a historical base rate, not a forecast.
For what it's worth: the multiplier-3 choice is already doing most of the filtering work. On the busiest lane (4H), roughly 4–5% of flips turn out to be flip-backs within their window – comparable-timeframe alternatives we studied reverse several times more often.
06 · Attention states
Attention states
An attention state is a label answering one narrow question: which assets changed recently enough to be worth a look right now? It is a filing system, not a forecast. Every rule below is mechanical – given the same candles, you can reproduce the classification by hand and get the same answer we do.
The four states are evaluated top-down as a ladder, and the first match wins: an asset carries at most one state, never two. If a fresh flip has already been invalidated, it shows as Invalidated and nothing else.
- S1 Invalidated
- Rule
- A 1D bull flip that was invalidated – price closed back through the Supertrend line frozen at the flip bar, or the lane re-flipped Bear, within 5 closed dailies of the flip – and that resolved within the last 7 days.
- Means
- Recent bull flip was invalidated within 5 closes – historically the worst outcomes.
- Evidence
- On the 1D primary lane, flips invalidated inside the 5-close window averaged −30.6% with a 0% win rate, against +15.9% for the roughly 98% that survived it. It is a rare event (about 2% of flips) and therefore a validated risk marker, not an edge – it is far too infrequent to move aggregate results. A separate study of 331 completed rides found the same asymmetry from the other end: rides that ended with a close back below their original flip level averaged −29%, versus +43% for the rest.
- S2 Fresh flip · 1W aligned
- Rule
- A 1D bull flip inside its confirmation window – within the last 5 closed dailies, still watching or already past the window without being invalidated – and that same asset's own 1W primary direction is Bull right now. Own weekly, current – not BTC's, and not whatever the weekly happened to be on the day of the flip.
- Means
- New daily bull flip while its own weekly trend is Bull – the only combination that improved results out-of-sample.
- Evidence
- A pre-registered multi-timeframe study locked five variants before running them. Exactly one cleared its kill criteria: restricting daily bull flips to those that occur while the asset's own weekly is Bull, and treating a weekly turn to Bear as the end of the hypothetical ride. Out-of-sample (2022 onward) it beat the same rules without the gate on both locked metrics — profit factor 2.00 vs 1.19, mean per-asset compounded return +44% vs −4% – and it was broad rather than concentrated: 15 of 28 assets positive, still +8% vs −29% with the top two removed. That is one out-of-sample window covering 92 trades. Read both numbers with the caveat below – they are a comparison between two research variants, not a return you can expect.
- S3 Fresh flip · counter-weekly
- Rule
- The same freshness as S2 – a 1D bull flip within the last 5 closed dailies, not invalidated – but the asset's own current 1W primary is Bear, or unknown (no weekly lane, or a lane too young to report).
- Means
- New daily bull flip against (or without) a Bull weekly – the weekly gate skips these; shown for completeness, not endorsement.
- Evidence
- This state is the exact complement of S2: the weekly gate skips roughly 70% of baseline flips, and these are the ones it skips. We label them rather than hide them – the same posture the flip watcher takes with counter-regime flips, and the same reason we publish flip-backs after the fact. No claim is made about them in either direction; the honest statement is that they fall outside the one condition that improved anything out-of-sample.
- S4 Near line
- Rule
- No flip has happened. The 1D lane's price simply sits within 5% of its current Supertrend line, or the still-forming daily candle is currently crossed to the other side of it.
- Means
- Price is within 5% of the daily line – a flip may confirm at an upcoming close. Distance says nothing about quality.
- Evidence
- Purely organizational – this is the scanner's "about to flip" list surfaced as a state, and the 5% threshold is a display cutoff, nothing more. It carries no claim because distance-based filters have been tested and rejected: a pre-registered study of flip-bar distance thresholds found the rule suppressed zero flips at every threshold tested (the Supertrend line jumps to the other side of price at a flip, so the close is always far from the new line), and the descriptive follow-up ran mildly the wrong way – closes that landed furthest past the level were followed by weaker outcomes, not stronger. Nearness is not a quality signal in either direction.
- — No state
Everything else – and that is most assets, most of the time. Absence is information, not an omission: an asset with no state is one where nothing mechanical happened recently, which is the normal condition. An empty attention list is a real answer, and it is the answer more often than not.
Scope
Attention states cover crypto only, on the 1D primary lane. Stocks, metals, FX, indices and on-chain DEX tokens carry no state at all, because the daily flip history and weekly context these rules read are only computed for crypto. That is a statement about our data coverage, not about those markets – a stock showing no state is not a stock where nothing is happening, it is a stock where we do not compute this. Inside the crypto 1D lane, read no state as "nothing here needs you." Outside it, read it as "not covered."
Read every number above with this attached
As a return engine, the flip strategy is not validated. Out-of-sample (2022 onward) the underlying rules earn roughly 0–7% a year – cash-like to BTC-like. The large full-period figures elsewhere in the research are artifacts: a single trade accounts for about 70% of all profit, one bull cycle dominates the sample, and the asset universe is survivorship-biased.
What is validated out-of-sample is drawdown control: every gated variant held its worst drawdown to −20…−29% over a stretch where BTC itself drew down −67%. And the weekly gate's result is relative – better than the same rules without the gate, on one out-of-sample window of 92 trades.
So "PF 2.00" above means "twice the profit factor of the ungated baseline, in one out-of-sample window" – not "twice as good" and not a rate of return. We do not quote it without this paragraph, and neither should anything that links here. Whether any of it is worth acting on rests on a belief the backtest cannot supply: that another broad crypto uptrend happens at all.
Organizational, not predictive
These states tell you where to look, not what will happen. None of them implies an action, and none of them is a recommendation to do anything. They exist so that a page of a few hundred assets has an obvious reading order on load.
There is deliberately no score, no rank and no risk number attached to any of them – within a state, rows are ordered by flip recency and nothing else. That is not modesty, it is the finding: the pre-registered 8,568-flip study described above found no feature that predicts, at flip time, which flips fail. A number here would be invented.
Some things are left out on purpose. Fresh bear flips are not an attention state – short outcomes are roughly flat-to-negative in the same research, and a bear flip is already visible as the end of a trend rather than the start of an opportunity. "Confirmed 3/3" is not its own state either: about 98% of bull flips survive the window, so the counter is bookkeeping and appears as a detail, never as a headline. And BTC's own regime stays page-level context – it is not a per-asset state, because the gate that survived out-of-sample was the asset's own weekly, not BTC's.
07 · Cycle position
Cycle coordinates, and the stage labels we tested and declined
"Where is this asset in its cycle" splits into two questions. One is descriptive and we answer it. The other asks for a name – young, late, topping – and we tested that one and don't publish it.
What ships is coordinates, each stated with the window it was measured over:
- Trend age percentile – how long the current ride has run, ranked against that asset's own completed rides on the same timeframe: "already longer than 82% of the 41 completed 1D rides on record." The current ride is unfinished, so it is never in the distribution it is compared against, and only rides that already ended are. "Already" is load-bearing: a ride at the 82nd percentile of finished rides is not 82% finished.
- Extension percentile – the distance from price to the Supertrend line, ranked against that same distance over the retained bars before it, with the bar being described excluded from its own distribution. It states the distance and the rank and stops; distance says nothing about quality, for the reason given under Near line above.
- Alignment – the 1D and 1W directions as a tuple, e.g. "1D Bull · 1W Bear", with no word attached to the combination.
Both percentiles are gated: trend age needs 10 flips on the timeframe, extension needs 120 retained bars. Under those floors the line renders nothing at all, because a percentile over a window too short to state is a window wearing a percentile's authority. The cycle ledger does the same thing one level up – real 10y rate, ratio regimes, the dollar's trend tuple, equity/alt/meme breadth percentiles, meme churn, funding extremes – every number with its window, ranking nothing.
The label we declined. The obvious next step is a Dalio-style stage word on top of those coordinates: young / mature / old for an aligned trend, "contested" when the weekly disagrees with the daily. We pre-registered that classifier and its pass criteria on 2026-08-07 and ran it the same day, over 321,328 point-in-time bars — 9,555 rides, 83 symbols, 55 years (1971–2026). All three criteria failed.
| Locked question | Bar, set beforehand | Measured |
|---|---|---|
| Does an old ride end sooner than a young one? | ≥ 10 pp, same sign in both halves | +7.04 pp |
| Does forward 10-bar drift differ by stage? | ≥ 1.82 pp best vs worst | +0.43 pp (sign inverts between halves in crypto) |
| Does "contested" add anything to the alignment tuple? | ≥ 10 pp hazard or ≥ 1.82 pp drift | +1.33 pp / −0.32 pp |
The age effect is real. Old aligned rides end within the next ten bars 30.5% of the time against 23.4% for young ones, monotone through 28.3% in the middle, same ordering in both halves of the sample and in both asset classes. It is also 3 points short of the bar we set before looking, and that is the finding rather than a technicality: a "late-stage" word implying 7 points on a base rate of roughly 25% would mislead more than it informs. The tilt it does describe is already in the age percentile that ships. No stage label ships anywhere.
That is the fifth pre-registered folk intuition to come back under its own bar on this data – the near-line distance filter (twice, and running the wrong way, above), a breadth-rotation alarm, macro-event direction labels, a jobs-day gold effect, and now cycle stages. We record them here for the same reason we publish flip-backs after the fact: "we checked" is worth something only with the number attached.
08 · Data hygiene
Staleness & data hygiene
Bad prices produce bad flips, so every lane is gated before it's trusted. For crypto, if an exchange candle's price diverges from Hyperliquid's own oracle price by more than 5%, that asset is flagged stale and any flip it would report is withheld – that combination usually means a thin or temporarily broken market, not a real move. For stocks, metals, FX, and indices, the equivalent gate is calendar and freshness awareness: no session, no candle, no flip.
The same closed-candle rule from earlier applies here too: a forming candle never triggers an alert, staleness or not. Alerts fire only on confirmed, closed-candle flips that pass the freshness gate for their market type.
09 · Disclaimer
What this is not
Arq is a research and informational tool. It is not financial advice, and it does not execute trades. Nothing on this site is a recommendation to buy, sell, or hold anything.
Any position sizing, entries, or "returns" you see referenced anywhere on the site are hypothetical, paper-only experiments for studying the indicator's behavior – no fees, no funding, no slippage, and no liquidation mechanics.