RSI Kernel Optimized, on MT4 and MT5, does not use 70 and 30. It goes through your chart, records the RSI reading at every confirmed swing high and every confirmed swing low, and builds a distribution out of them.

Then it tells you how close the current reading is to where this instrument's pivots have actually happened.

Arrows with the RSI value and the probability at each one, EURUSD H1

Look at the numbers under the arrows in that screenshot. 63.9. 61.0. 37.5. 35.1. Those are the RSI readings the tool considered pivot-like on that chart, and none of them would have triggered an overbought or oversold rule.

How the distribution is built

Swing highs and lows are found with a pivot needing 21 bars on each side. That's deliberately long: the point is to collect real turning points, not every wobble.

At each confirmed pivot the RSI value on that bar is stored. Highs and lows go into separate collections. Then a kernel density estimate is run over each collection, across 100 bins spanning the RSI range.

A kernel density estimate is a smoothed histogram. Instead of dropping each reading into a bucket, it spreads a small bump around it and adds the bumps together, so twenty pivots produce a smooth curve rather than a spiky bar chart. The kernel shape and the bandwidth control how much smoothing happens:

Setting What it changes
Kernel The bump's shape: Gaussian, Uniform or Sigmoid
Bandwidth How wide each bump is. Higher means smoother and less specific
Nº Bins Resolution of the curve, 100 by default
Probability Mode Nearest reads the density at the current RSI; Sum reads the whole tail beyond it

The peak of the high curve is the RSI value most associated with tops on your chart. It is what the tool has learned instead of 70.

The training split is real

Here is the design decision I'd install it for.

Learning data defaults to 80%. The first 80% of the bars the indicator can reach are used only to build the distributions. No arrows are drawn there. Signals begin after that point, on data that did not go into the model.

That's a train-and-test split inside an indicator, and it's uncommon enough to be worth stating plainly: most tools that claim to learn from your chart are scoring themselves on the same bars they learned from, which will always look good and means nothing.

The consequence you need to plan around: on a freshly loaded chart the arrows only appear in the last fifth of it. That is not a bug and there's no setting that fixes it without breaking the split. Load more history if you want more signalled bars.

What a signal actually is

Probability Mode decides the question being asked.

  • Nearest asks whether the current RSI sits close to the peak of the pivot distribution. Signals cluster tightly around one value.
  • Sum, the default, asks whether the cumulative density beyond the current reading is above a threshold. Signals appear across a broader band and start earlier in a move.

Activation Threshold is High, Medium or Low, and it moves the bar for both modes. Medium is the shipped default.

Every arrow carries two labels: the RSI reading, and the probability as a percentage. Both can be turned off. Keep them on for the first week; the numbers are what teach you what the tool thinks a pivot looks like on your instrument.

The panel in the corner shows the live RSI and the current bullish or bearish probability, so you can watch it climb rather than waiting for the arrow.

Where it's weak

Twenty-one bars either side means pivots confirm 21 bars late. The distribution is built from properly confirmed turning points, which is the right way to build it, and it means the newest pivot is never in the model yet.

The model rebuilds from the loaded history each time. Change timeframe, change a setting, restart, and the distributions are recomputed over whatever bars are available. Two charts of the same symbol with different history depths will disagree.

It's a distribution of RSI, not of outcomes. The percentage says "readings like this one have appeared at pivots on this chart", not "price is about to turn". A market that has changed character will produce confident signals from a distribution built on behaviour that has stopped happening.

The pivot collections are capped. Only the most recent readings are kept; older ones fall off the front. On a very long chart the model reflects the recent regime, which is usually what you want and is worth knowing.

One buffer, and it only holds the bullish signal flag. The bearish side is chart objects only, so iCustom gives you an incomplete picture. Use alerts if you need both sides out.

Alerts are off by default. Notification type ships as Do not notificate.

Questions you'll have

Does RSI Kernel Optimized repaint? Signals are placed on closed bars behind a new-bar check and stay where they were placed. The distribution behind them keeps updating as new pivots confirm, so a reading that qualified last month might not qualify today. Drawn arrows don't move.

Why are there no arrows on the left of my chart? By design. The first 80% of the loaded bars are the training window and produce no signals. Lower Learning data if you want signals earlier, and accept that they're then being scored on data the model saw.

What does the percentage under the arrow mean? How much of the pivot distribution the current RSI reading accounts for, under your probability mode. It's a similarity measure against past pivots on this chart, not a probability that price will turn.

Why don't the signals appear at 70 and 30? Because the tool doesn't use those numbers. It uses the RSI values at which pivots actually occurred on your instrument, which are usually much closer to the middle.

Which probability mode should I use? Sum, the default, for signals that appear earlier and more often. Nearest for fewer signals that sit tight around the peak of the distribution.

Best timeframe? H1 and H4, where a 21-bar pivot describes a real swing. On M5 it works but the pivots are noise, and the distribution it learns is a distribution of noise.

Can an EA read the signals? Only partly. There's one buffer, carrying the bullish flag. The bearish arrows are chart objects. Alerts carry both.

MT4? Both platforms are included, one download each. Same kernel options, same training split, same labels.

Support

Message me through the site with a screenshot, the symbol and the timeframe, and I can tell you what the distribution learned and why a bar did or didn't qualify. Updates are included for as long as your subscription runs.

Analysis tool. It places no trades and makes no profitability claim. Past price behaviour is not a guide to future price behaviour.