Argus Gaze: Attention Trend Engine TradingView Indicator

Written by Dominic Walsh · Published 22 Jul 2026

Argus Gaze works like a pattern recognition indicator TradingView traders can interrogate rather than trust blindly. Also, it encodes each bar as three ATR-normalised numbers, then compares today against the last 64 bars. A softmax turns those similarity distances into weights. Indeed, the engine averages the moves that followed the closest matches and smooths the result into a Trend Score. A prop firm risk layer then decides whether a signal may fire.

It is free and open source. Still, you can add it to any chart from the Argus Gaze: Attention Trend Engine script page on TradingView, and the full Pine source is published there for you to read.

What Argus Gaze: Attention Trend Engine plots on the price chart

Argus Gaze paints the 21-period EMA as a two-layer ribbon. Thus, it turns green while the Trend Score sits above zero and pink below it. Signals print as stacked triangles under or over the bar, tagged AGZ, and the signal bar itself changes colour. Hence, a soft background tint marks the active regime and turns orange during a lockout. Each signal draws a gold entry line plus dashed stop and target lines with price labels. Next, a twelve-row dashboard and a floating attention focus label complete the display.

How Argus Gaze: Attention Trend Engine is built

Three numbers per bar

The engine describes every bar with three values. Then, feature one is the one-bar close change divided by ATR(14). Feature two is bar range over ATR(14) minus one, so an average-sized bar reads zero. Yet, feature three is the volume z-score across the last 50 bars. All three carry no units. Truly, that matters, because the comparison that follows only works when a bar from 2019 and a bar from today sit on the same scale.

Now the attention step. For every lag k from 1 to 64 the script takes the squared distance between today’s three numbers and the numbers from k bars back. It divides that distance by the temperature input, default 1.0, then flips the sign. Closer states score higher. A softmax with max-subtraction converts the 64 scores into weights that add up to one.

Each past state carries a known sequel: the ATR-normalised move that came after it. Plainly, the engine averages those 64 sequels using the attention weights, then runs a 5-period EMA over the result to produce the Trend Score. Nothing here trains. Also, there are no learned parameters and no neural network. It is a weighted nearest-neighbour average wearing transformer vocabulary.

The gates a signal must clear

A raw trigger needs the Trend Score to cross above +0.08 or below -0.08, stated in ATR units. Indeed, it also needs conviction above the 0.05 floor. Conviction is one minus attention entropy divided by ln(64). Still, low entropy means a handful of analogues dominate the average. At temperature 1.0 the weights usually stay fairly flat, which explains why the default floor sits so low.

The risk engine rules next. Thus, presets load two limits: FTMO 5 percent daily and 10 percent total, Topstep 2 and 4, Apex 2.5 and 5. The script converts them into R multiples using your risk per trade. Hence, at 1 percent risk, FTMO becomes minus 5R for the day and 10R of drawdown. A simulated tracker then runs the signals at a 1.5 ATR stop and a 2.0 reward-to-risk target, checking the stop before the target on each bar.

Five more gates follow: the optional higher-timeframe EMA(21) filter, the session window, the news blackout, a cap of five signals per day, a five-bar cooldown, and barstate.isconfirmed. Next, the daily lockout always blocks signals and clears at the next session day. A drawdown breach only blocks when you switch that option on, otherwise it just flags DD WARN on the dashboard.

How to read the signals

Start with the dashboard. Then, trend Score states the attention-weighted expected move in ATR units, so 0.15 means the closest analogues moved roughly 0.15 ATR after states like this one. Conviction states how concentrated the attention was. Yet, a high score with 4 percent conviction rests on a broad, flat average. The same score at 40 percent rests on a few sharply similar bars.

The floating label tells you where the engine looks. Truly, it lists the top three lags, their weights and the effective number of analogues, computed as one over the sum of squared weights. When that figure sits near 64, attention spreads evenly and the forecast approaches a plain average. Plainly, when it falls into single digits, a handful of bars drive the whole reading.

Signals need a closed bar, so the triangle lands after the candle completes. Direction also alternates. The position state moves to long or short and never returns to flat on its own, so a buy can only be followed by a sell. Also, treat the gold entry line and its dashed pair as a suggested frame built from your ATR stop and reward-to-risk inputs, then size the trade yourself.

Every Argus Gaze: Attention Trend Engine setting explained

The script exposes 34 inputs, grouped in the settings panel exactly as shown below. Defaults are the published values.

Algorithm

Setting	What it does	Default	Range
Attention Lookback (N)	How many historical states the attention head compares against the current bar. Larger N = richer memory but slower regime adaptation.	64	16 to 200
Softmax Temperature	Controls attention sharpness. Low temperature = attention collapses onto the few closest analogues (spiky). High temperature = attention spreads across many states (smooth).	1.0	0.1 to 10.0
Score Smoothing (EMA)	EMA length applied to the raw attention forecast to produce the Trend Score.	5	1 to 50
Volume Z-Score Window	Window for the volume mean/stdev used in feature 3 of the state vector.	50	10 to 200

Signal Logic

Setting	What it does	Default	Range
Score Threshold	Trend Score must cross above +threshold (buy) or below -threshold (sell). Units are ATR-normalized expected return.	0.08	0.01 to 1.0
Conviction Floor	Minimum attention conviction (1 – normalized entropy) required for a signal. Keep low/permissive; raise to demand sharply focused attention.	0.05	0.0 to 0.9
Signal Cooldown Bars	Minimum bars between consecutive signals.	5	1 to 100
HTF Trend Confirm	Require price to be on the correct side of a higher-timeframe EMA(21). Uses the last CONFIRMED HTF value (lookahead off, offset 1) — non-repainting.	off	on / off
HTF Timeframe	Controls htf timeframe.	240

Prop Firm Rules

Setting	What it does	Default	Range
Prop Firm Preset	Loads daily-loss and max-drawdown limits for the selected firm. Apex uses a trailing drawdown. Custom reads the three inputs below.	FTMO	FTMO, Topstep, Apex, Custom
Account Size $	Sets the account size $ used in the calculation.	100000
Risk per Trade %	Percent of account risked per trade. Presets SUGGEST a value (FTMO 1.0 / Topstep 0.5 / Apex 0.5) but this input always rules.	1.0	0.1 to 5.0
Custom: Daily Loss %	Only used when Preset = Custom.	4.0	0.5 to 20.0
Custom: Max Drawdown %	Only used when Preset = Custom.	8.0	1.0 to 50.0
Custom: Suggested Risk %	Only used when Preset = Custom. Informational — shown on the dashboard.	1.0	0.1 to 5.0
Max Signals / Day	Consistency rule: no further signals after this many in one session day.	5	1 to 50
Lock Signals on Max DD Breach	OFF (default): a max-drawdown breach of the simulated R curve only shows DD BREACH on the dashboard. ON: it also suppresses signals until you reset the indicator. The daily-loss lockout always gates signals and auto-resets each day.	off	on / off

Sessions & News

Setting	What it does	Default	Range
Session Filter	Restrict signals to a trading session (exchange timezone of the chart symbol).	Off	Off, London, New York, RTH Futures, Custom
Custom Session	Controls custom session.	0800-1700
Enable News Blackout	Block signals inside the two news windows below (e.g. CPI 08:30, FOMC 14:00 ET).	off	on / off
News Window 1	Controls news window 1.	0825-0845
News Window 2	Controls news window 2.	1355-1415

Trade Management

Setting	What it does	Default	Range
Stop Loss (ATR mult)	SL distance = this multiple of ATR(14) from entry.	1.5	0.5 to 10.0
Reward : Risk	TP distance = SL distance x this ratio. Default 2.0 = 1:2 R:R.	2.0	0.5 to 10.0
Trade Line Length (bars)	Sets the trade line length (bars) used in the calculation.	20	5 to 200
Keep Last N Trade Drawings	Sets the keep last n trade drawings used in the calculation.	3	1 to 20

Visual

Setting	What it does	Default	Range
Show Dashboard	Toggles show dashboard on the chart.	on	on / off
Show 3-Layer Signal Glow	Toggles show 3-layer signal glow on the chart.	on	on / off
Show Trend Baseline Ribbon	Toggles show trend baseline ribbon on the chart.	on	on / off
Show Attention Focus Label	Floating label showing WHERE the attention head is looking: the top-3 most similar historical lags, their weights, and the effective sample size.	on	on / off
Buy Color	Colour used for buy color.	#4ade80
Sell Color	Colour used for sell color.	#f472b6
Accent (Cyan)	Colour used for accent (cyan).	#22d3ee
Entry Line (Gold)	Colour used for entry line (gold).	#fbbf24

Alerts built into Argus Gaze: Attention Trend Engine

The script ships 12 alert conditions. Indeed, open the alert dialog on the chart, pick the indicator as the condition source, then choose the event you want. Alerts fire on the close of the bar, so they follow the same confirmed-bar rule the on-chart signals use.

• AGZ Buy
• AGZ Sell
• AGZ Any Signal
• AGZ Regime Bull
• AGZ Regime Bear
• AGZ High Conviction
• AGZ Lockout
• AGZ Session Open
• AGZ Daily Reset
• AGZ Webhook JSON

Other markets and timeframes

Every input to the forecast stays relative. Still, the features divide by ATR and standardise volume, and the score threshold reads in ATR units, so the defaults carry to gold, indices and crypto without retuning. Two parts do not travel as cleanly. Thus, the volume z-score needs a volume feed, and the position size line uses syminfo.pointvalue, which behaves differently on futures than on spot forex. Session presets follow the exchange clock of the chart symbol, so check them per instrument.

Limitations worth knowing

This places no orders and manages no account. Hence, the stop, target and lot size on screen are suggestions built from your own inputs. The engine sees no spreads, no slippage and no news calendar. Next, the daily R figure and the win-loss count come from its own hypothetical trades at a fixed 1.5 ATR stop, so they describe the indicator, not your broker balance.

The analogue search has a narrow view. Then, it compares three features across 64 lags on the current chart only, so it cannot reach a similar setup from two years back. At lag 1 the outcome it borrows is the current bar’s own move, which already feeds feature one. Yet, attention entropy also stays high at the default temperature, so conviction rarely climbs far above the floor.

Volume drives the third feature. Truly, spot forex volume comes from broker tick counts rather than an exchange, so the z-score shifts between data providers. On symbols with no volume the feature flattens to zero and the state vector loses a dimension. Plainly, the prop firm presets are hard-coded numbers too, not a live rule feed, so check them against your firm’s current terms.

Get Argus Gaze: Attention Trend Engine on TradingView

Open Argus Gaze: Attention Trend Engine on TradingView

If you also trade MetaTrader, the MT4 and MT5 indicator library is available below.

Using it alongside MetaTrader

Adding a script on TradingView takes one click, so there is no install step here. Indeed, if you want the same idea on MetaTrader, the MT4 and MT5 indicator installation guide walks through copying files into the data folder and attaching them to a chart. You can also browse the full MetaTrader indicator library, the MT4 indicators section, or the other free TradingView scripts published on this profile. For related chart tools see the MT5 indicators section and the forex trading strategies guides.

External references

• Candlestick pattern on Wikipedia
• ADX at Investopedia
• Candlestick Bearish Reversal Patterns at StockCharts ChartSchool

Frequently asked questions

Does Argus Gaze repaint?

No. Still, both signal booleans require barstate.isconfirmed, and the plotshapes read those same booleans, so a triangle only prints after the bar closes. The optional higher-timeframe filter reads ema21[1] through request.security with lookahead off, so it uses a closed higher-timeframe bar. Thus, the dashboard and the attention label describe the live bar and update while it forms.

Is this really a transformer?

No. Hence, it borrows the softmax attention formula and nothing else. There is no training loop, no learned weight matrix and no network. Next, the query is today’s three features, the keys are the last 64 bar states, and the values are the moves that followed them. A fair description is a softmax-weighted nearest-neighbour forecast.

The score crossed the threshold. Why did no arrow print?

Seven gates can block it: conviction under the floor, the five-bar cooldown, the five-signals-per-day cap, the session window, the news blackout, an active daily lockout, or direction alternation after a same-side signal. Then, the dashboard shows Status, Signals Today and Session, so you can see which one stopped it.

Does it enforce my prop firm rules?

No, it simulates them. Yet, the daily loss lockout and the drawdown reading run on an internal R curve built from the indicator’s own signals at a fixed stop and target. Your firm’s dashboard remains the authority on your real limits. Truly, treat this panel as a discipline prompt, and confirm the preset numbers against your current account terms.

How much should I rely on this indicator?

Treat it as one input, not a decision. Plainly, it is a chart analysis tool, not trading advice. Test it on your own markets and timeframes before relying on it. Results are not guaranteed; past performance is not indicative of future results.