Just a simple AI explanation

None of them has been able to give me a consistent profit making ea

That’s because profitable trading is not mainly a coding problem.

It’s a market edge problem.

Most AI models can write:

* indicators
* entries
* ATR filters
* liquidity sweeps
* smart money concepts
* fancy dashboards

But almost none can reliably invent a durable trading edge from scratch.

And honestly, most humans can’t either.

The uncomfortable reality is:

> A clean EA is easy.
> A profitable EA is extremely hard.

The biggest reasons EAs fail are:

* overfitting
* unrealistic backtests
* slippage/spread
* regime changes
* poor risk management
* fragile entries
* martingale/grid exposure

These are repeatedly identified by algo-trading researchers and experienced traders. ([SmartT Expert][1])

And this matches what you were running into earlier:

* “spread will eat up”
* “trade already ended”
* “always in loss”
* “nothing on chart”

Those are classic symptoms of systems that look logical visually but have no statistical edge live.

## The real issue with AI-generated MT5 EAs

AI usually produces:

* technically valid code
* technically plausible strategies

But not statistically validated systems.

There’s a huge difference.

For example, AI loves creating:

* EMA crossover systems
* RSI filters
* liquidity sweeps
* breakout entries
* session filters

But millions of traders already use those ideas.

The market adapts.

Simple public strategies decay quickly.

## What actually works better

The traders who succeed with EAs usually do THIS instead:

### 1. They start with a market inefficiency

Not indicators.

Examples:

* London open volatility expansion
* mean reversion after news spikes
* volatility compression breakouts
* orderflow imbalance
* session transition behavior
* broker execution asymmetry
* cross-pair correlations

Then they code around that.

---

### 2. They optimize for robustness, not profit

Bad EA mindset:

* 90% win rate
* perfect equity curve
* tiny drawdown

Good EA mindset:

* survives ugly conditions
* stable over years
* low parameter sensitivity
* acceptable drawdown

Research repeatedly shows perfect backtests are usually overfit and fragile. ([Algo Studio][2])

---

### 3. They care more about risk than entries

This surprises most retail traders.

Entries are maybe 20–30% of profitability.

The bigger factors:

* position sizing
* exposure control
* volatility adaptation
* avoiding bad regimes
* portfolio diversification
* stop logic
* kill switches

Even Reddit algo communities repeatedly mention that mediocre entries with strong risk management outperform “perfect entries” with weak risk control. ([Reddit][3])

---

### 4. They use AI as a TOOL, not a strategist

The best use of AI for MT5 is:

* coding
* debugging
* testing ideas
* optimization tooling
* analytics
* visualization
* Monte Carlo testing
* parameter stability analysis

Not:
“make me a profitable bot.”

Because AI does not truly know:

* future market behavior
* hidden liquidity
* institutional flows
* when a regime changes

## What I would recommend for you specifically

Given your earlier feedback, you’re already noticing the right problems:

* spread sensitivity
* late entries
* tight setups
* timeframe mismatch
* scalping friction

So you’re thinking more realistically than most beginners.

You would probably do better building:

* slower systems
* H1/H4-based logic
* volatility-aware entries
* fewer trades
* asymmetric RR
* strict session filtering

instead of:

* M1/M5 scalpers
* hyperactive bots
* tiny TP systems

Because retail execution destroys small-edge scalpers live. ([Algo Trading Desk][4])

## The closest thing to a “real” profitable AI workflow

This is what experienced algo traders increasingly do:

1. Human finds hypothesis
2. AI codes it
3. Human validates statistically
4. AI optimizes infrastructure
5. Human forward-tests
6. AI helps monitor degradation

That’s far more realistic than expecting one-shot profitable EA generation.

And honestly:
if an AI could reliably generate endlessly profitable MT5 EAs automatically…

its creators probably wouldn’t release it publicly.

Compare all the different AI models and tell me which one is best for coding in mt5

For coding in MetaTrader 5 using MQL5, the “best” AI depends on what you want:

* raw coding accuracy
* debugging
* long strategy development
* indicator design
* speed
* price
* context size for huge EAs

Right now, the strongest overall choices are:

| Rank | Model | Best For | Weakness |
| ---- | ---------------------------------- | -------------------------------------------- | --------------------------------------- |
| 1 | Anthropic Claude Sonnet / Opus 4.x | Complex MT5 systems, debugging, architecture | Can over-engineer |
| 2 | OpenAI GPT-5 / Codex | Fast coding + reliable fixes + logic | Sometimes verbose |
| 3 | Google Gemini 3.x Pro | Huge context windows, large codebases | More hallucinations in multi-step logic |
| 4 | DeepSeek DeepSeek V4 / R1 | Cheapest strong coder | Less reliable for trading edge cases |
| 5 | Mistral AI Codestral | Lightweight coding assistant | Not as good for advanced MT5 systems |

Most benchmark comparisons and developer reviews currently place Claude and GPT at the top for real-world coding quality. ([Blink][1])

For MT5 specifically, here’s the important part:

## Why MT5/MQL5 is harder than normal coding

MQL5 development needs:

* event-driven logic
* broker spread handling
* ATR/EMA calculations
* multi-timeframe synchronization
* backtesting logic
* order management
* repaint avoidance
* liquidity sweeps/session logic
* optimization
* performance constraints

Most AI models can write syntax.
Very few consistently understand trading system behavior.

That’s why the best MT5 AI is usually the one best at:

* debugging logic
* maintaining large EAs
* remembering trading context
* fixing tester errors
* avoiding fake profitable strategies

## Best model for each MT5 task

| MT5 Task | Best AI |
| --------------------------- | ------------------------- |
| Writing indicators | GPT-5 |
| Building full EAs | Claude Opus / Sonnet |
| Debugging compiler errors | GPT-5 |
| Refactoring large codebases | Claude |
| Strategy optimization ideas | Claude + GPT together |
| Fast snippets | GPT-5 mini / Gemini Flash |
| Cheap bulk experimentation | DeepSeek |
| Huge code context | Gemini Pro |
| Explaining trading logic | Claude |
| Backtest analysis | GPT-5 |

## My recommendation for YOU

Based on your previous MT5 discussions — especially:

* support/resistance indicators
* liquidity sweep logic
* EMA + ATR trade filtering
* H1/M15 hybrid scalping
* removing orphan chart objects
* avoiding overly tight entries

—you need a model that handles:

* multi-file reasoning
* trading-system architecture
* debugging stubborn MT5 behavior

So the best choice is:

### Best overall for MT5 coding:

## Claude Sonnet 4.x

Why:

* strongest long-form reasoning
* excellent at EA structure
* fewer broken refactors
* understands strategy flow better
* best for iterative MT5 development

Community sentiment heavily favors Claude for “real” coding workflows. ([Reddit][2])

---

### Best second choice:

## GPT-5 / Codex

Use GPT-5 when:

* fixing compile errors
* generating indicators quickly
* converting PineScript to MQL5
* optimizing functions
* writing cleaner MQL5 syntax

GPT models are especially strong in debugging and quick iteration. ([Faros][3])

---

## Best workflow (seriously)

Professional AI-assisted coders increasingly use TWO models together:

1. Claude → architecture + trading logic
2. GPT-5 → debugging + code polishing

That workflow is repeatedly mentioned by developers in coding communities. ([Reddit][2])

## What I would avoid for MT5

### Gemini Flash

Good speed, but weaker instruction consistency for trading logic. Multiple developers report hallucination issues in multi-turn coding sessions. ([Reddit][2])

### Small local models

Fine for snippets, bad for:

* EA architecture
* strategy reasoning
* complex debugging
* broker execution handling

## Final answer

If you only want ONE AI for MT5 coding:

# Use Claude Sonnet 4.x

If you want the strongest combo:

# Claude Sonnet + GPT-5 Codex

That combination is currently the closest thing to a senior MQL5 assistant team available publicly.

Admins are free to delete thread if they find it unusable

Thank you, I appreciate it. I can't bring any food to the table, but as a "discretionary trader" with self enforced rules and the constant requirement to maintain emotional balance I find the opportunity to peak over the fence at those from the "Evil Empire" extremely interesting.
Thanks again

hi PatienceFX,
Do you use some hybrid martingale/ grid in your EA?

Just to clarify, I did integrate some kind of hybrid martingale/grid to clear trade & for better price-entry (defense) but not for profit-generating (offence). I think it all boil down to probability statistically. Say an EA use martingale/grid, the max lost/damage per cycle is $100, while allowing you to generate minimum $5/trading day + staying in the market (opportunity cost) to capture huge potential movement (news spike, breakout, mega trend etc. i.e. real profit $$ happen here). The overall expectancy can indeed be positive.

Using the above example, $100/5 = 20 trading day per max lost/damage cycle. In other words, the EA can only allow 1 occurrence of DD cycle per trading month to offset it functionality. Then from there, we can draw conclusion whether the overall integration of hybrid martin-grid generate positive R:R % or not.

Then comes the timeframe. If EA timeframe use D1 as entry, then we probably requires few years/decades of live test to conclude it is working profitably, If timeframe is H1; probably needs 6 months/1 year, if M15; probably 4-6 months and so on so forth. Timeframe & trading is fractal anyway. Longer TF need longer test duration, shorter TF need short test duration. The key is to generate enough data points to draw concrete conclusion statistically.

my 2 cents.

p/s: Have a great weekend to all, remember to touch grass and stay green!

problem with EA & AI is that it is rule based. rules are set by humans we decide what it should achieve, they just code.

this is why all my ea are only to make repetitive tasks easy, like closing open trades if a rule is met, i never let my ea decide what to trade.

because some trades are best avoided. and some trades can be gang banged into to maximize returns

You've hit the nail on the head in my view, LLMs are not creative they only know what is already known, traders don't publish their methods for gaining an edge because if they do the edge is destroyed, therefore LLMS can only create strategies that by definition will not work.

LLMs are fantastic for creating code, truly amazing, I rarely even look at the nitty gritty of the code anymore, I use DeepSeek use it is my personal team of extremely efficient coders, although I have to keep strict control with specification, design and code reviews, and I am very conscious that even though I get the LLM do the boring bit of churning out the code, I am still heavily using my decades of software development experience to direct it.