Understanding the consensus card

What the card is

The consensus card is the final conclusion for a pair. It sits under the widget header and, in full form, under the chart on the pair detail page. The header line shows the action badge, a one-line headline and the confidence percentage; click the line to collapse or expand the details.

Badge	Meaning
GO LONG / GO SHORT	A proven direction with enough agreement to act on.
LEAN LONG / LEAN SHORT	A majority direction that is just under the bar, or has a dissenting horizon, or a long during a crash watch. Not a fresh entry signal.
WAIT	No proven combination is firing, or the edge is too weak. The headline says why.
NO LONGS	Crash level is warning or critical: longs are blocked regardless of the votes.

An I took this button appears on LONG, SHORT and LEAN cards; it records the order you placed in your exchange app. See My Positions.

How the call is made

Every technique on every timeframe votes between -1 (short) and +1 (long): trend structure, moving-average ribbon, MACD momentum, RSI reversion, Bollinger position, order flow (CVD and taker ratio), volume-price divergence, distance from the 200 MA, the best chart pattern, the MA5/10/30/60 bundle breakout and the MA5/MA10 cross (see The MA5/10/30/60 bundle), on each of the nine timeframes; plus funding rate, open-interest build-up, long/short ratio, liquidation flow, order-book imbalance, the crash radar and Bitcoin's own 1h and 4h state.

During deep analysis the engine replays the pair's history every 30 minutes since 2018, records every vote, and checks the move on four horizons. It then searches for combinations of techniques (solo, pairs, triples and quads) that, whenever all members agreed, predicted the move. A combination is kept only if its accuracy beats coin-flip with a statistical margin on the first 75% of history and still does on the unseen last 25%.

Live, the card looks at which kept combinations are firing right now (all members agree on a direction). Only those count. Each contributes its edge above 50%, conditioned on the current regime (1h trend up, down or range) and trading session (Asia, Europe, US, weekend) when enough history exists. A combination that lost money in the current regime or session is suppressed.

The full card lists Firing combinations (proven on history), each with its direction, members, barrier win rate, conditioned win rate and the number of cases, and underneath its edge after costs, the typical drawdown before it pays (p50 / p80) and the median time to target (pays in ~2.5 h). Hover a line for the search and holdout accuracies. See How a set-up is judged a win below for what these mean.

How a set-up is judged a win

Until recently a past set-up was scored by where price stood at the end of the horizon: up at the deadline was a win, down was a loss. The path in between was ignored, so a set-up that first dived 3% and crawled back by the deadline counted as a win, while one that paid +1% quickly and then reversed counted as a loss. That is exactly the pattern of "favoured for a while, then a long way negative" that the engine ended up preferring. The scoring is now path-aware:

Triple barrier. For every historical sample the engine walks the 5-minute candles forward and places two barriers around the entry price: a target barrier at target barrier (5m ATR) above (default 2.0 x the 5-minute ATR) and a stop barrier at stop barrier (5m ATR) below (default 2.0). Both are multiplied by the square root of the horizon in 5-minute bars when widen barriers with the horizon is on, so the 4-hour horizon tolerates roughly seven times the wobble of the 15-minute one. The label is then:

What happened first	Label	Weight
Target barrier touched before the stop barrier	win	1.0
Stop barrier touched before the target barrier	loss	1.0
Neither touched by the deadline	sign of the move at the deadline (ignored inside ±0.3 ATR)	0.5

Both barriers in the same 5-minute bar count as a deadline outcome, because the order is unknown. A short set-up is scored on the mirrored path. The win rate, Wilson lower bound, holdout check and per-technique accuracy you see on the card all use these labels; the gradient-boosted combination model keeps the deadline return as its target. The stored definition is saved with the model (label_def) and old models are retrained by the next deep analysis.

Drawdown before it pays. For every winning sample the engine also records the deepest move against the trade before the target barrier was reached. Per combination the card shows the median and the 80th percentile of this drawdown (p50 / p80), in percent of price. This is the wobble you have to sit through on a typical winner, and it is what the default stop is placed beyond.

Edge after costs. Being right is not enough; the move must pay for the round trip. Per combination the engine computes the expectancy in ATR, win rate x median favourable move − loss rate x median adverse move (using the realised first-passage excursions), converts it to percent of price with the pair's median ATR, and subtracts two Bybit taker fees (2 x 0.055%) plus one funding settlement (0.01%): 0.12% in total. The result is shown as edge +0.31% after costs.

When a combination may fire. All three must hold: the Wilson lower bound of the barrier win rate is above 50% (as before), the edge after costs is above min edge after costs (%) (default 0.15%), and the 80th-percentile drawdown before it pays is at most max drawdown before pay (ATR) (default 3.0 x the 5-minute ATR). Combinations that pass the first test but fail one of the others appear on the card under Proven direction, not used · why not, each with its reason (no edge after costs (+0.04% per trade, bar 0.15%), too much drawdown before it pays (p80 3.8 ATR, bar 3.0)). They never move the call.

Pullback rule (do not buy the pop). When the direction is proven but price is more than max extension from EMA20 (ATR) (default 1.5) ATR above the 20-period EMA of the acting horizon's timeframe (15m, 1h, 4h or 1d), the card does not say LONG. It says wait for pullback and replaces the entry range with a Pullback zone: the EMA20 of that timeframe ± 0.3 of its ATR. The mirror applies to shorts. Independently, a call is held back as chasing when the last three bars of that timeframe already moved more than 2 ATR in the trade's favour. These vetoes are listed on the card in orange; the direction and the rest of the analysis are unchanged, only the entry is deferred.

All thresholds are in Settings > Analysis > How a set-up is judged a win (admin). Barriers change the labels and need a deep analysis to take effect; the edge and drawdown bars and the extension filter apply at once.

Horizons and conviction

The small chips 15m, 60m, 240m, 1440m show, per horizon, the direction, the number of combinations firing and the conviction: the agreement between firing combinations scaled by how many fire (three or more counts as full). The chip with a ring is the horizon the card acts on.

A horizon counts towards the decision when its conviction is at least 30%. The majority direction across horizons wins. If some horizons say long and others short the card reports horizons conflict and does not trade.

The acting horizon is chosen by Settings > Analysis > Trade duration (consensus horizon):

Setting	Effect
horizon preference = shortest acceptable horizon (default)	The shortest horizon whose conviction is at least horizon margin (default 0.75) times the best one, so you get 1 to 4 hour trades unless only the daily horizon has conviction.
horizon preference = strongest conviction	Always the horizon with the highest conviction; tends to favour 1440 minutes.
max horizon (min)	Horizons longer than this are never suggested (60, 240, 720 or 1440).

Confidence is the historical accuracy of the best firing combination, damped when few combinations fire. GO LONG or GO SHORT requires: trained history, confidence at or above the threshold, at least one combination firing, conviction of at least 50%, at least two horizons agreeing, and no conflict. A direction that fails one of these but is within 8 points of the threshold becomes LEAN.

Reading a WAIT

A WAIT headline is explicit, for example WAIT, long edge too weak: confidence 64% below the 70% bar, only 1/4 horizons agree. The possible reasons are: confidence below the bar, fewer than two horizons agree, or firing combinations disagree. no proven combination is firing on any horizon means nothing with a proven edge is active; no combination with an edge after costs is firing (3 proven direction(s) rejected) means set-ups fired but history says they do not pay after costs or wobble too much first; the why-not list names them. wait for pullback (long): entry zone 1h EMA20 ± 0.3 ATR means the direction is proven but price is extended: the Pullback zone box tells you where to wait for it. history not analysed yet (plain vote average) · run deep analysis means the pair is not prepared; do not act on it.

Crash gating

• Crash level warning or critical turns any long call into NO LONGS.
• Crash level watch caps a GO LONG to LEAN LONG and adds crash watch active: long capped to LEAN to the headline. The engine never issues a fresh GO LONG while a crash watch is active.
• Short calls are not blocked by the radar. When the crash-short backtest says shorting at the current risk level paid on this pair, the card switches to a CRASH SHORT headline; see Crash short call and backtest.

Other lines on the card

Exit: names the exit policy that paid best on this pair; macro event: flags a high-impact release inside the hour; pattern engine: quotes what the pattern engine alone says (it is already one of the techniques). The note under the card explains the ranges and any stop widening from the analog forecast.

Tip: the Techniques & combinations tab on the pair detail page shows every kept combination per horizon with its holdout accuracy, so you can judge how much evidence sits behind a call.