Perceptrader AI EA Review (2026): What a 'Neural Network Grid' Is, and What Its Own Signals Show

44.0 / 100

weighted verdict score on eight published criteria

1.51

profit factor of the vendor's main PAI signal over 6,755 trades, with an average loss larger than the average win

$2,800

Market price on 21 Oct 2026, purchase only, last updated 8 Feb 2026

The verdict at a glance: a grid with an AI label, and signals that do not isolate the AI

Perceptrader AI earns points for showing live accounts and for a documented, configurable design. It loses them on the central question for a product with 'AI' in its name: does the AI do anything the grid underneath would not? The signals do not answer that, the risk controls are those of a grid, and the price is the highest on the Market. The scorecard and radar show how the total was built.

What Perceptrader AI actually is

Perceptrader AI MT5 is an Expert Advisor for MetaTrader 5, sold on the MQL5 Market by Valeriia Mishchenko of Russia, whose seller profile shows a 4.2 rating from 277 votes, 14 products and 11 signals. It was published on 24 July 2023, showed version 2.43 and was last updated on 8 February 2026. It trades NZDUSD, USDCAD, AUDNZD, AUDCAD, NZDCAD and GBPCHF from one M5 chart. The listing calls it a grid system 'using deep learning and artificial neural networks'.

Read closely, the design has three layers. The foundation is a grid, with a configurable distance, a 'smart distance' that adapts to volatility, custom lot multipliers, a weighted take profit, break-even after a chosen grid level and a trailing stop. The second layer is a set of trend, momentum and volatility filters 'based on deep learning'. The third is a 'Perceptron-based minimum profit probability' setting: the EA estimates the probability that a trade will profit and skips it below a threshold. There are two trading modes, 'Standard' and 'Unique', and optional forecast filters that call ChatGPT and Gemini, which the listing says work 'in live trading only'.

How we judged it, and who has something to sell

Evidence on EAs is never neutral. Here is every party near this product and what each stands to gain, ours included.

• The developer sells the EA, sells signal subscriptions (999 USD a month each) and runs the live accounts. Her records are public, which counts for something, and they are also her sales page.
• MQL5 and MetaQuotes host the Market, the reviews and the signals. The statistics and warnings on each signal are computed by the platform.
• Buyers who wrote reviews are few: seven reviews in three years. Testimony, not data, and each is unverifiable from outside.
• Reseller and affiliate sites sell or review the product with figures that conflict with each other. We use none of them as evidence.
• ForexCopilotEA sells a listing for this EA. We gain if you buy it, which is why this verdict opens with that fact and applies the same rules as every other.

The vendor's PAI signals: what they show, and what they cannot

The main signal is 'PAI Unique Mode No Limits with AI Filter' (signal 1919885): an MT4 account at IC Markets with 1:300 leverage, 5,000 USD initial deposit, 8,000 USD withdrawn and no deposits. It showed a balance and equity of 11,216.13 USD and growth of 347.44% on 21 October. The platform lists its start as 8 April 2023 and its length as 377 weeks, which would be about seven years; the author's label reads 'growth since 2019'. The week count, the start date and the label cannot all describe the same history, so we treat the record as covering the product's logic, not as 377 weeks of this exact version.

When a signal's own numbers cannot all be right

The second signal, 'PAI Standard Mode 5 Pairs with AI Filter' (2020263), shows a balance of 1,611.67 USD from a 400 USD start, yet a growth figure of 0.61%. It shows a profit factor of 33.25, a best trade of 56,646.03 USD on that 1,612 USD balance and a recovery factor of 683. Its own page says 'Trading style has changed. Part of history is not included in statistics.' Those figures describe an account whose history is partly excluded and whose units cannot all be in the same currency. We cannot tell which numbers are right, so we do not use any of them as evidence of profitability, and we advise buyers to ignore a profit factor above 10 on any signal, here or elsewhere. A figure that good is a data problem before it is a strategy.

The seller's own list adds a third AI-filter signal, 'PAI Unique Mode 5 Pairs' (2020261), which shows +291% over 178 weeks with a 73% drawdown and 1,320 trades. A fourth, 'Waka Clone', shows 0% growth with a 90% drawdown, and a fifth, 'WakaSmart', shows a profit factor of 73.76 with 57% drawdown. They are different settings of related systems, not independent confirmations, and the spread of results (17% to 73% drawdown, profit factors from 1.5 to 275) is the honest picture of what 'Perceptrader AI' can look like depending on the mode and risk used.

A note on the figures. MQL5 pages are generated from the account, but one page can show the same measure two ways: the main signal prints both 1.51 and 1.50 for profit factor and gives two readings of its longest winning streak. We quote the main statistics block and name the basis, and where a page disagrees with itself we say so.

What 'AI' means here, and what it does not

The listing's AI claims fall into three groups. The first is a set of machine-learning filters, described as deep-learning models for trend, momentum and volatility, and a perceptron that estimates a minimum probability of profit. A perceptron is a very small neural network, and a threshold on its output is a legitimate filter. It reduces the number of trades; it does not change what a grid does when price keeps moving against it. The second is a note that the neural network 'has been retrained using the latest historical data', which means the model is periodically refit, a real maintenance burden and a risk, because refitting on recent history is how an over-fitted model is produced. The third is the optional ChatGPT and Gemini forecast filters, which work 'in live trading only'.

The honest test of an AI filter is a controlled comparison: the same grid, same pairs, same period, with the filter on and off, over a long window and out of sample. The signals names ('with AI Filter') imply the comparison exists somewhere. It is not on the listing, and the vendor's signals do not carry an 'off' twin. Without it, a buyer is paying $2,800 for a grid and trusting that the AI adds value. Our walk-forward guide and backtest versus live guide show how to test it yourself.

A grid with multipliers: what that means for risk

Underneath the filters, Perceptrader AI is a grid. If price moves against a first position by a set distance, the EA opens a second, then a third and more, each with a lot multiplier, and the basket closes at a weighted take profit. The listing's risk controls are the right ones for that structure: a maximum floating drawdown that closes everything and pauses trading, daily drawdown limits like a prop firm's, a minimum free margin and a cap on open lots. The default 500-pip stop, used only if the setting is left at zero, is wide, and the controls are only as good as the values you choose.

A filter that skips low-probability entries makes a bad first trade less likely. It cannot stop a good-looking first trade from going wrong, and the grid's exposure when that happens is the same with or without AI. That is why we cap the risk score by design family and why the 17% to 73% drawdown range matters more than any single signal. See our grid, martingale and scalper comparison.

What buyers say, and what reviews cannot tell you

The listing shows a rating of 4.67 from just seven reviews over three years, with 30 comments. Six are positive and short: one reports consistent profit for over a year and praises the number of adjustable parameters (April 2026), one calls it a 'stable safe grid' (February 2024), one cites a solid track record (January 2024), one says it works well, especially the AI filter, after two months (October 2023), and two are a bare rating and a vote of faith in the author from launch week. One review, in August 2023, calls support slow and the price poor value; the developer replied that she found no message from the buyer, that she usually answers within a day and gave a support address.

Seven reviews is a thin base, and the rating says little. The same seller's older product, Waka Waka MT5, has 44 reviews and a lower rating, with several detailed negative accounts of deep drawdowns; see our Waka Waka verdict. Perceptrader AI inherits the grid's character, and its small review count is a gap, not a clean bill.

The real price of Perceptrader AI

$2,800, purchase only, is the same price as the vendor's other EAs, from a decade-old grid to this AI version, and the listing was last updated in February 2026. That uniform price says the price is set by the seller's brand, not by the evidence for each product. The only mitigation is a free demo with 6,929 downloads that runs in the tester. Add a VPS and the capital a grid needs, which at the vendor's own guidance is thousands of dollars.

Can your account actually run it?

The setup is demanding. It needs a hedging account, a VPS, six pairs from a single M5 chart, correct broker suffixes, and access to three external websites for news and time data. Backtests are to be run with Tick Data Suite using GMT+2 with US daylight saving. The leverage guidance is the key sentence: at 1:30, low-risk settings only on accounts under $6,000. See our ESMA guide, our hedging versus netting guide and how much money to run an EA.

On prop firms: the listing includes FTMO-style daily drawdown limits, which is thoughtful. A grid that adds positions into a losing basket is still a poor fit for most prop-firm rules, and an EA that depends on web access to three third-party sites needs the firm's permission. See our FTMO guide.

Will it last?

The vendor is long-standing, with a family of grid EAs running since 2021 or earlier, a website and a support address. Against that, the listing was last updated in February 2026, eight months before our check, which is a long gap for a product that describes a continually retrained neural network. The grid underneath depends on currency pairs that tend to range, and trends in AUDCAD or NZDCAD are where reviewers of the sister product report their worst losses. The AI filter may help or may be an overfit to the training window; the data does not say.

What we could not see

Criterion by criterion: how the 44.0 was built

Each score below is out of 10, multiplied by the criterion's weight, and sits inside the tier described by the written scoring anchors. The reasoning is short on purpose; the sections above carry the evidence.

Who it might suit, and who should walk away

• It might suit an experienced trader with several thousand dollars of risk capital and a hedging ECN account, who will test the filter on and off before trusting it, and who sets the maximum-drawdown input before the first trade.
• It might suit someone who wants a configurable grid with prop-style risk limits and accepts that the AI label is unproven.
• Walk away if you are buying because of 'AI' or '80 consecutive months'. Neither is something the record isolates.
• Walk away if your account is under $1,000, or under $6,000 on 1:30 leverage and you want more than the lowest risk setting.
• Walk away if you cannot sit through a 30% to 70% drawdown, or cannot afford to lose the account.

If you want to test it: a 12-week plan

1. 1

   Step 1: Weeks 1 to 2: read the signals

   Open signals 1919885, 2020263 and 2020261. Note which mode each uses, the drawdown, and anything on the pages that disagrees with itself.

2. 2

   Step 2: Weeks 2 to 4: run the filter on and off in the tester

   Use Tick Data Suite, the same six pairs and the same period, with the profit-probability filter on and off. Record the profit factor, drawdown and trade count for each.

3. 3

   Step 3: Weeks 4 to 10: forward-test on a demo

   Same version, lowest risk preset, the maximum-drawdown input set. Record every basket's depth and whether the filter skipped trades.

4. 4

   Step 4: Weeks 10 to 12: decide with numbers

   Continue only if the filter-on results beat filter-off with a margin, and the deepest basket is one your account survives.

5. 5

   Step 5: Start small, and only if it passed

   Fund an amount you can lose completely and treat the first live month as part of the test.

Ten questions to put to the developer before you pay

A developer with a real edge can answer these in a paragraph. Vague or missing answers are information too, and they are the fastest way to close the gaps this verdict could not.

1. Is there a controlled comparison of the same grid with the AI filter on and off, and can I see it?
2. Which version and settings does each PAI signal run, and are the ChatGPT and Gemini filters on?
3. How can the 377-week count, the 8 April 2023 start and 'growth since 2019' all be true for signal 1919885?
4. What explains the profit factor of 33.25 and the best trade of 56,646 USD on signal 2020263?
5. How is the '80 consecutive months in profit' counted, and for which account and product?
6. What is the deepest basket in the history of the main signal, and when?
7. When was the neural network last retrained, and on what data window?
8. Why has the listing not been updated since February 2026?
9. What are the default multipliers and the default maximum-drawdown setting?
10. What is the refund policy, and what happens to a buyer whose account is stopped out?

What would move this score

1. A published filter-on versus filter-off comparison over several years. This lifts trust and profitability together.
2. An MT5 signal on the current version, of at least 12 months, with consistent statistics. This lifts evidence.
3. A resolution of the signal inconsistencies. This lifts trust and evidence.
4. An update and a stress test of each risk preset. This lifts sustainability and risk, though a grid will not leave the High label.
5. Downward risks: a deeper drawdown, a reset record, or a retraining that changes behaviour without notice.

The developer, or any reader with the signals open, can send corrections through the contact page. We will recalculate and log the change with the date at the top of this page.

How it compares with the other verdicts

Next to its sister product Waka Waka, Perceptrader AI shares the pair family and the grid, adds a filter and a six-pair set, and has a shorter, less consistent record. Next to Quantum Queen X at 36.5 and Wave Rider at 43.0, it has more evidence than either, with a deeper and more variable drawdown. Its 44.0 lands where a long vendor-run grid record with unresolved data issues tends to land. For how these system types differ, see our strategy comparison and our grid category.