Field guide

How to read a signal report

Every EtherSignals report, whether on Telegram, in the live feed or as a desktop alert, is the output of the same AI trend model. This guide explains where each number comes from and, more importantly, how to read it honestly. Five minutes here will save you from the most common misreadings.

1. Anatomy of a report

A report is a full snapshot of the model's view right now, so you never need earlier reports to interpret it. A typical Telegram push looks like this (illustrative values):

Telegram push · illustrative@ethersignals

1[ETHUSDT] 5m regime update · price 2,440.542▼ Strong Down  (score −0.510)32m −0.40  ·  5m −0.51  ·  30m −0.30  ·  2h +0.10  ·  1d +0.254Continuation (Win%): 58%5Patterns on close: MACD death cross 52%6R 2,489.7 (+2.0%) · PP 2,402.3 · S 2,405.7620d range 2,405.7 – 2,778.0 · BREAKOUT: inside

1Header: market (here ETHUSDT), the timeframe that triggered this report, and the current price.
2Signal line: the trend regime on the triggering timeframe, with its composite score. This is the headline.
3Five-timeframe board: the regime score on every tracked timeframe (2m, 5m, 30m, 2h, 1d). This is where confluence shows.
4Win%: how often this exact state continued historically. It is the most misread number in the report (see §4).
5Patterns: events that fired on the candle that just closed, each with its own hit-rate.
6Key levels: objective price levels, namely the pivot set and the 20-day Donchian range, plus whether price has broken out.

2. Trend regimes

The model classifies every timeframe into one of five regimes using the composite score (§3) plus a trend-strength gate (ADX):

Regime	Icon	Condition
↑↑ Strong Up	↑↑	score ≥ +0.42 and ADX ≥ 20
↑ Up	↑	score ≥ +0.16
→ Range	→	score between −0.16 and +0.16
↓ Down	↓	score ≤ −0.16
↓↓ Strong Down	↓↓	score ≤ −0.42 and ADX ≥ 20

You will notice that regimes never flip-flicker, and that is by design. Two dampers keep signals calm:

Hysteresis band. Once a trend is established, it takes a narrower score move (past ±0.08 rather than ±0.16) to drop back to Range, so trends survive minor pullbacks.
Confirmation candles. On the 2m and 5m charts a new regime must hold for two consecutive closed candles before it is confirmed (30m/2h/1d confirm on one).

Closed candles only

All analysis runs on candles that have finished printing. The forming candle never influences a signal, so reports never repaint.

3. The model score

The score is a weighted average of nine indicator "votes", each ranging from −1 (bearish) to +1 (bullish):

score = Σ(weight × vote) / Σ(weight)   ∈ [−1, +1]

Indicator	Weight	What it looks at
EMA stack	1.4	Price vs EMA20/50/200 alignment
ADX / DMI	1.4	Trend direction × strength (weak trends barely vote)
MACD	1.2	Histogram momentum, ATR-normalised
SuperTrend	1.2	SuperTrend(10,3) direction, full vote on fresh flips
LinReg drift	1.1	20-bar least-squares slope
Volume flow	0.8	Buy vs sell volume over the last 10 bars
RSI	0.7	Momentum position, capped when overbought/oversold
Bollinger %B	0.6	Position within the bands, discounted outside them
KDJ	0.6	Stochastic K−D cross

Read the magnitude, not just the sign: |score| beyond ±0.4 marks conviction; values loitering near zero mean the timeframe genuinely has no edge right now. Weights are not fixed (see §7), and they are clamped so no single indicator can dominate.

4. Win%, read it right

Definition. The model replays roughly the last 300 closed candles. Each time the same regime appeared on this timeframe, it checked: did price actually move more than 0.5×ATR in the regime's direction within the next 3 candles? The decay-weighted share of hits (recent history counts more) is the Win%.

It is not a next-move probability

`1d ↓↓ Strong Down 36%` does not mean "36% chance of falling tomorrow". It means: historically, when this exact state appeared, 36% of the time the decline continued past the threshold within 3 candles. It is a base rate for the state, not a forecast for the next bar.

Three reading habits that keep you honest:

Below 50% is normal. The denominator includes "no directional move within 3 candles" noise samples (~23%). In ranging markets every regime's Win% sinks together, which is the model honestly reporting that momentum has no edge.
Compare longitudinally. Strong Down climbing from 25% → 45% tells you recent down-moves are following through more often. The trend of the number matters more than the level.
Multi-timeframe confluence beats single rows. The 2m chart is close to noise by nature; the model's real value is Strong-tier filtering and alignment across 2m→1d.

Why you sometimes see “n/a”

If a state has fewer than 15 historical samples, the model refuses to print a number and shows n/a. A missing figure is honesty, not a bug. The same rule applies to pattern probabilities.

5. Patterns

Patterns are events that fired on the candle that just closed; they appear and disappear as candles roll. Each carries its own historical hit-rate under the same 3-candle, 0.5×ATR rule.

Pattern	Trigger	Direction
EMA bullish alignment	EMA20 > EMA50 > EMA200	Bullish
EMA bearish alignment	EMA20 < EMA50 < EMA200	Bearish
MACD golden cross	Histogram crosses above zero	Bullish
MACD death cross	Histogram crosses below zero	Bearish
SuperTrend flips bullish	SuperTrend flips to long	Bullish
SuperTrend flips bearish	SuperTrend flips to short	Bearish
RSI recovery	Previous bar < 30, current ≥ 30	Bullish
RSI fading	Previous bar > 70, current ≤ 70	Bearish
Donchian breakout	Close > highest of prior 20 bars	Bullish
Donchian breakdown	Close < lowest of prior 20 bars	Bearish
Bollinger squeeze	Bandwidth in lowest 15% of last 120 bars	Neutral (bets on expansion, not direction)

The squeeze is different

`bb_squeeze` is direction-neutral: it flags compressed volatility that tends to resolve in a large move (hit = any move beyond 1.2×ATR within 3 candles). Treat it as "get ready", not "which way".

6. Key levels

Levels come from two objective, formula-based methods, and nothing is fitted or hand-picked:

Classic pivots: computed from the previous completed trading day's High/Low/Close, PP = (H+L+C)/3, with R1 to R3 and S1 to S3 derived from it. Best suited to intraday reference.
Donchian-20: the highest high and lowest low of the last 20 closed daily candles. This is the key gate for day-level breakouts.

How to read the block:

Resistance / Support: the nearest merged level above or below the current price. The percentage in parentheses is the distance to it.
BREAKOUT / BREAKDOWN: price has moved beyond the 20-day high or low (a daily-level breakout in progress). inside means price is still within the range.

Pro tip

Pivot levels and Donchian levels that coincide at the same price form the strongest confluence zones. And during the model's warm-up on a new market the levels block may be absent entirely, because daily history simply isn't long enough yet.

7. How the model improves itself

The nine indicator weights are not hand-tuned constants. Every trend call is later graded against what price actually did, and the weights learn from the outcomes at two speeds: a small online update per graded call, and a fuller replay-based adjustment every 15 minutes. It is a multiplicative "expert" scheme: indicators that were right get louder, wrong ones get quieter.

Six guardrails keep the learning scientific rather than lucky:

Shrinkage prior: after every update, weights are pulled a little back toward their baseline, so they can't drift into over-fitted corners.
Hard clamps: every weight stays within [0.3, 2.2], so no indicator can monopolise the score.
Normalisation: average weight returns to 1.0 each round, so scores stay comparable across months.
Sample thresholds: batch updates need at least 120 decisive samples per indicator, otherwise they are skipped.
Time decay: statistics weight recent history exponentially more (half-life about 250 candles).
Version library with rollback: every weight change is archived as a version with its own live track record. If the current version underperforms a historical one decisively, the model rolls itself back automatically.

8. Reading this site's feed

The live feed shows the same snapshot, structured. Field mapping:

On this site	Report field	Meaning
Time	ts	When the model published the snapshot (shown in your local time)
Market	symbol + marketType	Pair and venue type (Perp or Spot)
Price	price	Latest price at push time
AI signal	timeframe + regime + score	The triggering timeframe, its regime, and composite score
2m → 1d board	signals[]	All five timeframes' scores; the ring marks the trigger
Continuation	prob	Win% from §4, shown as “n/a” when samples are thin
Patterns	patterns[]	Events on the last closed candle (§5)
Key levels	levels	Resistance / PP / Support / 20-day range / breakout (§6)
As published on Telegram	message	The exact channel text, verbatim

Open any row in the live feed to see the full breakdown for that signal.

9. Honest limitations

We'd rather publish the caveats than hide them:

Samples correlate. Consecutive candles are highly related, so treat a stated sample of 100 more like 30 when judging confidence.
Fitted to one market. The model learns on ETHUSDT perpetuals; behaviour on other symbols is not calibrated.
The 2m chart is noise-limited. Direction on ultra-short timeframes is near-random; the value there is Strong-tier filtering and multi-timeframe confluence.
Win% is history, not destiny. Markets change distribution, which is exactly why the statistics decay over time.
And the standing one: EtherSignals is a research and educational project. Nothing here is financial advice, and signals can be plainly wrong. Never trade on them.

Questions about a specific field? The live feed exposes the same data via the public feed API, and the community discusses reports daily on Telegram.