Is “Machine Learning RSI” Actually Tradeable? Settings, Signals and Caveats Explained

What this indicator is actually reading

In one sentence, Machine Learning RSI is an oscillator that digs up past bars where RSI was behaving the way it is right now, then puts what happened next to a vote.

A standard RSI is read through a fixed rule: above 70 is overbought, below 30 is oversold. Machine Learning RSI throws that framing out and cares far more about how RSI is behaving than about the number it prints. How fast is it moving, is that move accelerating, how extreme is it versus the last 100 bars, how far apart are the fast and slow RSIs — all of that gets bundled into a “momentum fingerprint” for the current bar, and the engine goes looking for past bars with a similar fingerprint.

Each historical match then votes: “price went up after this one,” “this one rolled over.” The tally becomes the current directional bias, and only after that bias clears two separate scores — Rank (setup quality) and Confidence (how sure the model is) — does a triangle marker finally print on the chart.

Sorting out everything it draws on your chart

Once you add Machine Learning RSI, you get visuals in both the price pane and the lower pane. It’s a lot to take in at first, so it helps to split it out piece by piece.

Where it appears	What you see	What it means
Lower pane	ML RSI line	A 0–100 line: standard RSI with a conviction-based tilt, then smoothed
Lower pane	Signal line (SMA 14 by default)	A smoothed reference line over the ML RSI. Crosses work as a secondary trigger
Lower pane	70 / 50 / 30 levels plus a tinted band	The overbought, neutral and oversold reference zones
Price chart	ML Supertrend and its glow	A trailing stop whose band width changes with model conviction
Price chart	Trend cloud	A four-layer gradient fill between the stop line and price
Price chart	Candle coloring	Regime color matching the Supertrend direction (default: uptrend blue, downtrend white)
Price chart	Triangle markers	Long (up) and short (down) entry cues that cleared every condition
Price chart	Small dots	The bar where the Supertrend flipped direction

Here’s what that looks like on a chart. The key point: candles, cloud and stop line all take their color from the same thing — the direction of the stop. Wherever you look, the regime color agrees, so color alone tells you which side the market is on.

DiagramImage

What Machine Learning RSI draws on your chart

The upper pane shows the regime and the trade cues; the lower pane shows an RSI adjusted by model conviction.

Vowars DE ver.3.11.0

The default palette is a little unusual: blue for up, white for down. If you’re used to green and red, recoloring is probably the first thing you’ll want to do. The triangle markers are the exception — they ship as lime for longs and red for shorts, so those read clearly out of the box.

Under the hood: eight RSI features, matched against history

1. Building eight features out of one RSI

Machine Learning RSI takes the base RSI (default 14) and computes these eight measurements on every bar.

• RSI value — the raw reading itself
• RSI slope — how much it has moved versus 3 bars ago
• RSI acceleration — the change in that slope: is momentum igniting or fading
• Distance from neutral (50) — which way it leans, and by how much
• RSI percentile rank — how extreme the current reading is within the last 100 bars
• RSI volatility — how erratic RSI itself is (steady momentum versus jumpy momentum)
• Fast/slow RSI spread — the gap between the half-length and double-length RSIs
• RSI regime — how far a 20-period smoothing of RSI sits from 50

The percentile rank is the one worth calling out. Plenty of instruments rarely tag 70 or 30 at all, but because the engine also asks “is this extreme by this market’s own recent standards,” it isn’t chained to the classic 70/30 levels. That design fits price action like Bitcoin’s, where RSI can stay pinned high through an entire leg up.

2. Banking history and labeling the outcome

On every closed bar, those eight features get stored alongside “what price did over the next 4 bars.” The cap on how many rows are kept is Memory Depth (bars) (default 500).

The “what price did” part is measured against ATR. Using Learning Sensitivity (×ATR) (default 0.5) as the yardstick, a close 4 bars later that’s within 0.5× ATR is labeled +1, up to 1× ATR +2, and beyond that +3 — giving a graded label from −3 to +3. It’s not a crude green-candle/red-candle split; each move is weighted by whether it was meaningful relative to volatility at the time.

DiagramImage

Every banked bar is labeled by how far price moved 4 bars later

Each stored bar gets a label from −3 to +3 depending on whether the close 4 bars later cleared 0.5× or 1× ATR.

Vowars DE ver.3.11.0

These labels become the weight of each bar’s vote in the next step. A +3 bar casts a bullish vote three times as strong as a +1 bar.

3. Finding the closest matches and letting them vote

The engine measures the gap between the current bar’s features and each stored bar’s features, then keeps the closest Analog Count (k) (default 8). It compresses those gaps logarithmically instead of using them raw, so one wildly mismatched feature won’t blow up the similarity read on its own. Outlier-resistant, in other words.

Those 8 matches each vote, and closer ones carry more weight. A positive tally means bullish, negative means bearish. When the tally lands in the middle, it’s treated as no direction at all, and no signal can come from it.

4. Learning which features matter, automatically

With Auto-Optimize Weights (on by default), the engine uses the banked history to ask, on every bar, “which feature separates the bullish outcomes from the bearish ones most cleanly?” and reassigns the weights accordingly. Features doing the heavy lifting get weighted up; the rest lose influence. No feature ever drops to zero, though — there’s a floor built in.

The optimizer needs at least 60 banked rows before it kicks in, so you won’t see true behavior right after loading a chart, or on symbols with thin history. More on that in the caveats below.

5. Filtering through Rank and Confidence

A direction alone prints nothing. This next part is arguably the heart of Machine Learning RSI.

Score	What it measures	What feeds it
Rank	Setup quality	Agreement among the matches, how tightly clustered they are, alignment with trend direction, healthy volatility, regime fit, RSI slope fit, and how long the current call has held. Penalties are then subtracted for chop, stretched conditions and rapid recent flips
Confidence	Model conviction	Mostly agreement and tightness of the matches, plus credit for how long the call has persisted and slope fit. Heavy deductions if the engine has been flipping repeatedly

By default, nothing prints unless Rank is 60 or higher and Confidence is 50 or higher. One practical detail before you touch anything: Rank realistically tops out around 95 and Confidence around 90. Assume they’re true 100-point scales and crank the gate to 80, and you’re asking for far more than you think. The default 60 is effectively demanding close to 70% of the achievable score.

The other easy-to-miss detail: on the bar where the call flips, the persistence credit is zero. Since signals only fire on that flip bar, Confidence has to clear 50 on agreement and tightness alone. In my runs, most of the skipped signals were caught right here on Confidence.

The ML Supertrend: a trailing stop that breathes with conviction

The line in the price pane looks like an ordinary Supertrend, but the band width isn’t fixed.

• High conviction — bands tighten, the stop trails closer to price, and trend flips are picked up sooner
• Low conviction, or chop detected — bands widen to stay out of the way of whipsaw

On defaults (ATR Multiplier 1.5, ML Band Adaptivity 1), the effective multiplier swings roughly between 1.5× and 3.0×. When conviction is near zero you’re effectively running a loose 3.0× stop, and framing it that way makes the behavior easier to read. The ATR length here is hard-set at 10.

DiagramImage

Low conviction pushes the stop away; high conviction pulls it in

The trailing stop never sits tighter than 1.5× ATR and widens up to 3.0× as conviction fades.

Vowars DE ver.3.11.0

As ② shows, conviction works regardless of direction. In a downtrend, strong bearish conviction pulls the stop in just the same. When the stop is flat, though, its distance to price is just carried over from where it last moved. If you want to read the current width, look at bars where the stop is actually moving.

This line isn’t just decoration. Through Trend Gate (Signals must align with trailing stop), its direction determines whether a signal can fire at all: longs only while it reads up, shorts only while it reads down. Candle coloring follows the same direction.

Putting it to work on a Bitcoin chart

1Leave it on defaults and read the regime first

Drop it on the Bitcoin 1H and the first thing to do is look at the candle colors and the direction of the Supertrend — not the wiggles in the ML RSI. You just want to know which side the market is on. When a trend is running, the stop tracks alongside price and the cloud stays thin. With no direction, the line drifts away from price and the cloud fattens up, so you can eyeball a range before analyzing anything.

2Check where momentum sits on the ML RSI

The lower-pane line is standard RSI with a conviction-based tilt added, then smoothed over 3 bars. At full conviction that tilt can push the line up or down by about 18 points, which is why it can reach 70 or 30 earlier than a plain RSI. If a plain RSI prints 64 while this one shows 73 on the same bar, that’s the tilt doing its job, not a bug.

DiagramImage

ML RSI drifts away from standard RSI in the direction of conviction

The dashed line is a standard RSI(14). Bullish conviction lifts ML RSI, bearish conviction drags it down, and the result is smoothed over 3 bars.

Vowars DE ver.3.11.0

As ② shows, the smoothing also means it lags a standard RSI on sharp pullbacks. An ML RSI break below 70 can come a little after the plain RSI’s.

Reading it is straightforward: above 50 favors the bulls, below 50 the bears, and pushes beyond 70 or under 30 mark strong momentum. A cross of the signal line (SMA 14 by default) works as an early heads-up that a full signal may be coming.

3Watch a few weeks of signals before touching anything

This is the step people skip, and it matters most. Every filter is on by default, so on something like Bitcoin that alternates sharply between trend and range, the marker count gets thinned out considerably. Spend a few weeks watching how often signals appear on your timeframe, decide whether that’s too many or too few, and only then start adjusting. Jump straight to moving thresholds and you’ll have no idea what caused what.

4Set alerts so you’re not glued to the screen

Machine Learning RSI ships with six alert conditions: long signal, short signal, any signal, trend flip up, trend flip down, and any trend flip. There’s also a message format that includes the ticker, the timeframe, and the Rank and Confidence at the moment it fired — so you can judge setup quality straight from the notification. A Rank 90 alert and a Rank 61 alert deserve very different treatment, even when both say long.

Every gate a signal has to clear

Getting a single triangle on the chart takes a surprising number of boxes ticked. In order:

1. The analog vote clears the neutral zone and commits to one side
2. Trend Gate (Signals must align with trailing stop) — direction must agree with the Supertrend (when enabled)
3. Volatility Band — ATR rank sits at or above the floor (20 by default) and at or below the 85 ceiling (when enabled)
4. Chop Filter — the bar isn’t flagged as range-bound (when enabled)
5. With all of the above satisfied, the call flips to the opposite side
6. Rank clears its threshold, and Confidence clears its own
7. At least 5 bars have passed since the previous signal
8. The bar is closed

The common misconception: a Supertrend flip and a triangle marker are two different things. The stop can flip up without a ▲ if the model’s call doesn’t change, and a ▲ can show up a few bars after the flip.

DiagramImage

A stop flip doesn't guarantee a triangle

▲▼ only print when the model's call flips and that bar also clears the Rank and Confidence thresholds.

Vowars DE ver.3.11.0

Bar ③ shows a particularly important behavior. If a long signal is skipped for lack of Confidence on the flip bar, the call itself stays switched to long, so no ▲ follows in that direction even once Confidence recovers. The next ▲ only comes after the call flips back to short and then to long again. As the figure shows, that can mean sitting out an entire move.

Settings in detail, and what to run

The learning engine

Signals and filters

Auto-optimization and feature weights

Supertrend and the RSI signal line

Four setups for four different jobs

Riding Bitcoin trends without overthinking it

Leave Trend Gate (Signals must align with trailing stop), Volatility Band and Chop Filter all on, and only drop Adaptation Speed to around 0.2. Keep the Rank and Confidence gates at their defaults. That narrows signals to moments where you’re with the regime, there’s energy in the market, and it isn’t ranging. It pairs well with Bitcoin on the 1H to 4H.

When you only want the A-plus setups

Push Min Rank to Signal to 65–70, Min Confidence to Signal to around 60, and lift Min Vol Rank to about 30. Raise Learning Sensitivity (×ATR) to roughly 0.8 as well, and the engine only learns from moves that actually went somewhere. Run this and you’ll have weeks with zero signals — fine, as long as you can sit on your hands.

When you want to miss fewer moves

If the “skipped on the flip bar, then sat out the whole move” pattern bothers you, one option is dropping Min Confidence to Signal to around 45. The Rank gate stays where it is, so low-quality setups still don’t get through. But you’ll also let in more split votes, so don’t take every ▲ at face value — use the Rank and Confidence in the alert to prioritize.

Fading the edges of a range

Switch Trend Gate (Signals must align with trailing stop) off and Chop Filter off, and you’ll start getting counter-trend signals inside ranges. Understand what that means: you’re deliberately unlocking the exact conditions this tool handles worst. Only run it alongside something that identifies the edges of the range — horizontal levels, volume profile, whatever you use. On its own, it will chop you up.

What I liked, and what to watch out for

Caveats, and how not to use it

On repainting

The entry triangles are built to fire only on closed bars, so you won’t see a signal appear and then vanish mid-candle. That part you can trust.

That said, the ML RSI line and the Supertrend line do update with price while the current bar is forming. That’s the same as a standard RSI or Supertrend, and the usual rule applies: don’t treat an unconfirmed value as a decision.

No lookahead, but performance is context-dependent

The training labels are built from “the 4 bars that followed an already-closed past bar.” So it isn’t peeking at future price to generate signals — that much is clear from the design.

At the same time, as the author openly acknowledges, this isn’t a trained neural network — it searches for similar precedents within the chart and classifies from them. So results shift substantially with symbol, timeframe, memory depth and filter settings. Nothing about the AI label makes it right more often.

Don’t trade Machine Learning RSI in isolation

Rank and Confidence quantify how well the historical precedents line up. They aren’t win rates and they aren’t expectancy. A Rank 85 signal failing is completely ordinary. Stop placement and position sizing are yours to define, separately from this tool. Using the Supertrend line as a stop reference is perfectly sound, but size accordingly: low conviction means a wider stop, and a wider stop means more risk per unit.

The best pairing: a simple horizontal line

What Machine Learning RSI doesn’t know is where price actually sits. The analog vote looks only at the shape of momentum, not at how much room there is to the nearest high or low. That’s exactly why tools that mark key price levels complement it so well — and the simplest, most effective one is a horizontal line at the recent high or low.

DiagramImage

Check where ▼ printed against a recent-low horizontal line

Triangles don't know where price sits. If support is right below, there isn't much room left to the downside.

Vowars DE ver.3.11.0

The workflow is simple: when a triangle prints, check the distance to the nearest level in the direction of the signal. As in the figure, if the ▼ bar closes only about 0.5 ATR above support, there’s little room left to the downside — so you skip it, or wait for a clean break of the level. Other pairings worth considering:

• Volume profile — finer-grained than a horizontal line, marking high-volume price areas as levels
• Higher-timeframe trend — with a 4-bar learning horizon, the read skews short-term by nature. Simply checking the daily or 4H direction and skipping signals that fight it changes how the tool feels
• Volume tools — the volatility filter is ATR-based, so volume itself is never looked at. Adding something that gauges whether a breakout has real participation fills that gap

Who this is actually for

Type of trader	Fit	Why
Values signal quality over signal count	Excellent	The double gate plus the filters make it selective by design
Primarily trend-following	Excellent	The trend-direction gate and the Supertrend already lean that way out of the box
Enjoys dialing in settings	Excellent	Plenty of inputs, and they genuinely change how it behaves
Wants to trade the arrows mechanically	Mixed	Without understanding the scores, it’s easy to lower the gates and gut the edge — and skipped signals mean missed moves
Trades mean reversion inside ranges	Mixed	That runs against the design, and means switching filters off
Keeps charts minimal	Poor	Line, cloud, candle coloring and markers — it’s a lot of visual weight

An AI RSI that makes silence work for you

Plenty of indicators claim “AI” and “machine learning,” and plenty turn out to be a couple of moving averages in a trench coat. Machine Learning RSI isn’t one of them: it genuinely banks past states, searches for similar ones, has them vote, and adjusts which features matter. The name and the contents line up, and that deserves credit.

Running it on Bitcoin, what I appreciated most was how much of the time it says nothing at all. The Chop Filter mutes it through ranges, and the volatility band shuts it down when things go quiet. If you’re tired of indicators that spray arrows across the screen, that silence is the feature.

On the complaints side: the sheer number of settings, the default colors, and the way a skipped signal leaves that direction quiet for a while. Adaptation Speed in particular swaps out the learned weights every single bar at its default value, so I’d drop it the moment you install.

Taken as a whole — a heavily filtered, high-spec RSI bundled with an adaptive Supertrend — this is a well-built piece of work. Internalize two things — Rank and Confidence don’t really run to 100, and Confidence struggles on flip bars — and tuning the thresholds stops being guesswork. It’s not a self-contained holy grail, but it adds one genuinely useful question to your discretionary read: what did momentum like this lead to last time?