ML RSI Classification takes the current state of RSI, finds the eight moments in recent history that most resemble it, and asks what happened next each time.
The vote becomes a direction and two scores. Then a trailing band on price uses those scores to decide how much room to give the trade.
Eight numbers describe the moment
RSI on its own is one number and it throws away everything about how it got there. This builds a description with eight:
the RSI value, its slope, its acceleration, its distance from the midpoint, its percentile against its own recent range, its volatility, the spread between a fast and a slow RSI, and a regime term.
That eight-number description is computed for the current bar and for every bar in a rolling memory of up to Memory Depth, 500 by default. Each historical entry also carries a label: what price actually did after it.
Finding the analogues
For the current bar it measures distance to every stored moment and keeps the Analog Count k closest, eight by default.
The distance is a Lorentzian measure rather than a straight Euclidean one, which is the same choice made in the other classifier in this range and for the same reason: it compresses the effect of a single wildly different feature. Two moments that match well on seven features and badly on one stay close together, where Euclidean distance would push them apart.
Each of the eight features carries a weight in that distance, and each weight is an input. With Auto-Optimize Weights on, which is the default, they're adjusted as the chart runs rather than left where you put them.
The eight neighbours then vote, weighted by closeness: a near analogue counts more than a distant one. If the weighted vote leans far enough either way, that's the direction.
Two scores, two gates
This is the part I'd point at.
The vote produces Rank and Confidence, and they measure different things. Rank is about how strong and well-formed the setup is. Confidence is about how much the neighbours agreed and how close they actually were.
A signal needs both: Min Rank to Signal at 60 and Min Confidence to Signal at 50. Passing one is not enough.
Both scores are also penalised when the search found fewer than k usable neighbours. If the current state is genuinely unlike anything in memory, the tool doesn't quietly extrapolate from three loose matches; it marks itself down for having little to go on. That is the honest behaviour and it's rarer than it should be.
The trailing band widens when the model is unsure
Here's the feature worth the money.
The ML Supertrend is an ATR band, but its multiplier is not fixed. It's the base multiplier scaled by how convinced the classifier currently is: strong agreement and close analogues pull the band in, weak or contradictory ones push it out. A separate chop filter cuts the drive further when the fast and slow averages are too close for the market to be trending at all.
So the band tightens when the model has a clear read and backs off when it doesn't. Look at the gold chart above: tight dots through the climb, then noticeably wider spacing across the sideways stretch at the top.
Almost every adaptive stop on the market widens on volatility. This one widens on uncertainty. Those are different signals and they part company exactly when it matters, which is a quiet market the model can't read.
| Ordinary RSI tool | This | |
|---|---|---|
| What describes the moment | One RSI value | Eight features including slope, acceleration and regime |
| Where the reading comes from | A fixed threshold | The eight closest historical analogues and what followed them |
| When history has no match | Signals anyway | Both scores are penalised |
| Trailing band | Fixed or volatility-scaled | Scaled by how convinced the classifier is |
| Feature weights | None | Eight, optionally self-adjusting |
What you get
Five plotted series: the ML RSI line, its signal average, a Bollinger pair around that average, and the signal arrows. The trail and the markers are chart objects.
Signal Line Type takes None, SMA, SMA+BB, EMA, SMMA or WMA, so the smoothing on the classifier output is yours to choose.
Alerts are wired to the buy and sell signals with separate switches, and Notification type ships as none, so pick a channel before expecting anything.
Where it's weak
"ML" here means k-nearest-neighbours, not a trained network. There is no training run, no model file, no gradient anything. It stores recent states and looks up the closest ones. That is a legitimate and well-understood technique, and it is not what most people picture when they read AI. I'd rather you buy it knowing which one it is.
The memory is this chart, and only this chart. Five hundred bars of the symbol and timeframe you're on. Nothing is shared across symbols, nothing is saved between sessions, and restarting the terminal means it rebuilds the bank from whatever history is loaded. Two charts of the same symbol with different history depths will not agree.
Neighbours are sampled, not exhaustive. The search steps through the bank rather than testing every stored bar, which keeps it fast and means the eight it finds are the eight best of a subset.
Auto-weights make results non-reproducible. With the weights adapting as the chart runs, the same bar can score differently after more history has loaded. Turn Auto-Optimize Weights off if you need a fixed, repeatable configuration.
The object count is high. Thirty-five thousand chart objects on gold H1 in this build, more on H4. The trail is drawn as individual dots. Two copies on one chart is noticeable.
Two gates interact. Raising either Min Rank or Min Confidence quiets the tool quickly, and because they measure different things it is not obvious which one is doing the filtering. Change one at a time.
The newest bars can't be labelled. A stored moment needs to know what came after it, so the most recent bars contribute to the picture only once enough has happened.
Questions you'll have
Does ML RSI Classification repaint? The classifier line and the band move with the forming bar, as any oscillator and any trailing stop do. Arrows are placed on closed bars. The memory bank keeps growing, so a signal is not recomputed after the fact but the model behind future signals keeps changing.
Is this really machine learning? It's k-nearest-neighbours with a Lorentzian distance and adaptive feature weights. A real technique, honestly named, and not a neural network.
Why do I get so few signals? Both gates have to pass, rank at 60 and confidence at 50, and both are marked down when the search finds fewer than eight usable analogues. Lower one at a time.
Why does the trailing band keep changing width? By design. The multiplier scales with how convinced the classifier is, so an uncertain read gives the trade more room. The chop filter widens it further in a flat market.
What are the eight weights for? They set how much each RSI feature counts when measuring the distance between two moments. Leave Auto-Optimize Weights on unless you want a fixed configuration.
Why do two charts of the same symbol disagree? The memory is built from the history each chart has loaded. Different depth, different bank, different neighbours.
Can an EA read it? Partly. The classifier line, its signal average and the Bollinger pair are plotted buffers. The trail dots and the arrows are chart objects.
My terminal has slowed down. Turn off Show ML Supertrend or Show Signal Markers. The trail is thousands of individual objects.
Best timeframe? H1 and H4. The memory is 500 bars, so on M5 that's under two days of context and the analogues are all from the same session.
MT4? Both platforms are included, one download each. Same eight features, same k, same adaptive band.
Support
Send me a message with a screenshot, the symbol and the timeframe, and I'll tell you what the tool was doing on that bar.
Analysis tool. It places no trades and makes no profitability claim. Past price behaviour is not a guide to future price behaviour.