Olivier Nomblot / Perfil

# AI IN MQL5: WHERE THE MARKETING ENDS AND THE ENGINEERING BEGINS

Everybody is selling an "AI EA" now. Open the MQL5 Market on any given day and half the listings have a neural network in the screenshot and the word "self-learning" in the title. Most of it is noise. Before you buy one — or before you build one — you need to understand what the language can actually do, because the gap between what people claim and what MQL5 permits is wide, and that gap is where buyers get burned.

I write these systems for a living. So let me be blunt about the limits.

## "AI" MEANS TRAINING. MQL5 WAS NOT BUILT TO TRAIN.

When people say AI, they mean machine learning: you take a large dataset, define a model, and run it through thousands of optimisation passes — gradient descent, backpropagation, the whole loop — until the weights converge. That is *training*. It is heavy, it is iterative, and it wants a real numerical stack: tensor libraries, automatic differentiation, ideally a GPU.

MQL5 has none of that as a first-class citizen. It is a single-threaded, event-driven language built to execute a strategy tick by tick, not to grind a training loop over millions of samples. There is no native autodiff. There is no PyTorch. There is no `model.fit()`. If you want to train a deep model the way the Python world trains one, MQL5 is the wrong room to be standing in.

So when a listing says "the EA trains a neural network live on your account," ask yourself what that actually means in a language with these constraints. Usually it means one of three things: it's marketing, it's a trivial linear update dressed up as a brain, or — the honest version — it's something far more disciplined than a buzzword, which I'll get to.

## YES, YOU CAN RUN ONNX. NO, THAT DOESN'T MEAN YOU CAN TRAIN.

Credit where it's due: MetaTrader 5 added native ONNX support, and it changed the conversation. You can now train a model in Python — PyTorch, TensorFlow, scikit-learn, whatever you like — export it to the ONNX format, drop the `.onnx` file in the terminal, and run *inference* inside a standalone EA at near-C++ speed. That is real, and it's a genuine step forward.

But read that carefully. ONNX runs a model you *already trained somewhere else*. The training still happens in Python, on your machine, offline. The EA is just executing a frozen brain. And the moment your frozen brain is wrong about the current regime, it stays wrong until you retrain it in Python and redeploy. There is no live adaptation in that loop.

It gets worse in practice. An ONNX model is a self-sufficient blob — it stores the network and nothing else. Feature engineering, scaling, dimension reduction, the entire preprocessing pipeline that fed it in Python: none of that ships inside the file. You have to reimplement all of it in MQL5, by hand, exactly, or the model receives garbage at the input layer. And here's the killer for anyone selling on the Market: a model optimised against your broker's data can behave completely differently on a customer's broker — different spreads, different feed, different fills. One ONNX brain across many brokers is a reproducibility minefield. I've seen products get hammered with bad reviews for exactly this reason.

## THE PYTHON BRIDGE EXISTS. ALMOST NOBODY SHIPS IT.

The other route people mention is the Python bridge — keep MT5 talking to a live Python process, do the heavy lifting there, send signals back. It works on your desk. I have used it.

In *production*? Forget it. You are now running two processes that have to stay alive in lockstep on a VPS. You need the right Python version, the right packages, the right environment, all surviving reboots and broker updates. You add latency on every decision. And you cannot ship any of it on the MQL5 Market — the Market sells a compiled `.ex5`, not a Python environment with a dependency tree. So the bridge is a fine research tool and a terrible deployment story. The number of commercial EAs actually running a live Python bridge on a client VPS is, in my experience, close to zero. People talk about it. They don't run it.

## THE REAL ALTERNATIVE: LEARN NATIVELY, LEARN ONLINE

So if you can't train deep nets in the language, ONNX only runs a frozen brain you have to babysit, and the Python bridge can't ship — what's left?

What's left is the approach I took with **A Black Box Pentagon Cortex**: build the learning *into* the EA, natively, and make it learn *online* — trade by trade, in real time, with no external process and no offline retraining cycle. Not deep learning for its own sake. Adaptive statistical intelligence that runs at native speed and ships as a single file.

That's the part the marketing crowd skips, because it's hard.

## WHAT PENTAGON CORTEX ACTUALLY DOES

Pentagon is a five-engine ensemble. Each engine has its own magic number, its own logic, and its own opinion on the market:

1. **The neural brain** — Gaussian-neuron network for the core signal.
2. **A Renko engine** — structure and noise filtering off brick action.
3. **An adaptive RSI / range engine** — exponentially-weighted RSI percentile bands that move with the market instead of sitting on fixed 70/30 lines.
4. **A momentum engine.**
5. **An RSI-SAR engine** for trend confirmation.

Five independent reads, combined. No single point of failure when one regime turns.

The neural brain is where the actual intelligence lives, and it is not a toy linear update. It carries a real statistical stack that runs entirely in MQL5:

- **GMM regime detection** — the EA classifies *what kind of market it's in* before it decides how to act. Trending, ranging, chaotic — different regimes, different behaviour.
- **A Kalman filter** to track the underlying state through noise.
- **Bayesian win-rate estimation** — it maintains a live, probabilistic estimate of its own edge instead of pretending the last hundred trades are gospel.
- **KDE modelling of the P&L distribution** — it understands the *shape* of its outcomes, not just the average.
- **Mahalanobis novelty scoring** — when the market looks nothing like anything it has seen, it knows, and it can pull back instead of charging into the unknown.
- **Monte Carlo simulation and walk-forward validation** baked in, so the thing isn't just curve-fit to one history.

And critically: **it persists its brain.** The learned state is written to disk with an integrity footer and rotating backups, saved per trade with periodic autosaves, and guarded so a corrupted brain auto-restores from backup. Reboot the VPS, the EA wakes up knowing what it knew yesterday. That is the difference between a system that *learns* and a system that *resets every time you blink.*

## WHY THIS IS CUTTING EDGE

Not because it has the word "neural" in it. Anyone can write that word.

It's cutting edge because it solves the deployment problem nobody else solves honestly. It learns live, online, with no offline retraining cycle to babysit. It has zero external dependencies — no Python process, no bridge, no fragile environment to keep alive on a VPS. It runs at native speed because it *is* native. It ships as a single compiled file on the MQL5 Market, which means it actually reaches the people who buy it and behaves the same way on their broker as it does on mine. And the v2.00 MT5 build came through Market validation on a 648-trade report — it doesn't just look good in a screenshot, it survives the validator.

That's the whole point. The frontier in this space isn't "who can say AI the loudest." It's *who can make a genuinely adaptive system that actually deploys, actually persists, and actually runs where the broker lives.*

## BOTTOM LINE

Be careful what you believe when someone sells you AI in MQL5. The language cannot train deep models. ONNX runs a frozen brain you have to retrain elsewhere and rebuild the pipeline for by hand. The Python bridge can't ship. Those are the facts, not opinions.

The thing that works is native, online, self-contained adaptive intelligence — built into the EA, persisting across restarts, running where your money runs. That's what Pentagon Cortex is. Everything else is a screenshot.

You don’t have to worry about back test especially when you have a five core neural. It’s right there on the screen WIN factor .PNL confidence factor PF factor back testing is unsure the data is unsure and it’s always curve fitted so it always looks good ; so that’s in the past go with the future all of my EA have the data right in front of you. You have neuraks that learn eternally , make decision trees , 5 agents working with one boss . Join the future , things change almost day by day and I prefer the latest in learning .

Think big

I ve got the best neurals available on mt4 ,mt5, self Learning , multi core take a leap before everyone else don t be a dinasaur.

Iif you are looking to join and be accepted in a prop firm, take it from an expert at a very low price for now ; Set files includrd or as for custom ones

NEW TODAY :I love my new EA Pentagon the Renko makes synthetic inner boxes and cAUght the big move down on Gold today while the 5 engine neural filters out bad trades learns from its mistakes and saves its weightings Every aspect of the market is covered at all times , even when you sleep .This is the best . i don;t know how you can trade without it .This is not amateur hour . Made by a pro trader with 30 years experiece in AAA banks .

WELCOME TO PENTAGON

"Some traders run one strategy. This runs five — and won't tell you which one just fired."

ALSO COMING SOON Five engines. One brain. Decades of edge, sealed inside one self-learning system.Guess who's back: five adaptive engines, one Cortex, decades of trading distilled.One body, five heads — the black-box system I built for myself.Pre-tuned, self-learning, five-engine powerhouse. Set your risk and watch it breathe.Decades of edge, five coordinated engines, one sealed brain. Welcome to ......

NEW TODAY