Mad Turtle EA MT5 XAUUSD Dedicated ONNX Deployment Machine Learning Half-Year Live Trading Profit 141%

Transaction Specifications

Recommended Broker: Tickmill / JustMarkets

Symbol: XAUUSD

Timeframe: H1-M15 (any)

Account type: Hedging account.

Leverage Ratio : 1:500

This EA is currently quite popular in the forex community, mainly because the product claims to use a machine learning model and also has relatively appealing live trading signals. Many people don’t really understand what a machine learning model is, but their curiosity and fascination have driven the EA’s strong sales. The backtesting is extremely slow, likely due to the multi-model combination algorithm. Looking at the live trading history, no anomalies were found. Interested friends can conduct an in-depth backtest themselves.

Key Project Information

Mad Turtle EA MT5 This EA does not connect to any GPT website or similar platform. The model is deployed using the MT5 built-in ONNX library (Open Neural Network Exchange format).

Deposit Safety: No pre-rolled strategies, micro-spread arbitrage, or narrow-range trades based on low statistical samples. Does not use grid or Martingale strategies, or any other dangerous strategies that could wipe out profits or deposits in a single day.

Uniqueness: The foundation is my unique feature engineering, using a custom training method based on rewards and penalties, cascading filtering through meta-models, and various ensemble methods.

Models are trained for the complete 24-hour trading pattern on large H4 timeframes, with holding periods from a few hours to several days.
The profits generated are clean and consistent, calculated in pips rather than lot sizes.

Signal Statistics:

Old WT Signal: 1,250 trades, producing 4.4 million pips, averaging ~3,500 pips per trade.

New ICM Signal: 103 trades, producing 127k pips, averaging ~1,200 pips per trade.

Key points from these statistics:

Low deposit load because large positions are not required.
Strong loss resilience, as the profit/loss ratio is roughly balanced. To lose accumulated profits using the same lot size, one would need to experience losses equal to many days of normal volatility. The more profit earned, the less likely losses are—a unique feature of this strategy. Comparing these stats with other signals shows how this EA differs from others.

Philosophy and Goals

The main goal of the Mad Turtle Project is to create a tool that generates profits more interesting than bank interest, without putting heavy burdens or deposit risks on the trader, and without disturbing your sleep.

No deposit-promotion games or millionaire dreams—just a long-term strategy suitable for serious traders who understand the value of safe trading.
The project is continuously evolving. Members can influence updates and share ideas, with good suggestions being implemented.

Not Suitable For

This product is not recommended for users who:

Expect fast profits and want to turn $500 into $50,000 quickly.
Frequently interfere with the EA’s trading, breaking profit/loss balance and blaming the EA.
Manually modify model coefficients or thresholds, causing losses and blaming the EA.

> This EA is designed for fully automated mode, which is recommended. The internal model and setting list is intended for experienced traders to test personally, not for live accounts. Only use developer-recommended settings.

Signals

Usage Notes

This EA is built on multi-class models, each containing 2–4 sub-models: some for buy signals, others for sell signals (2×2).

Models support each other but are not mirror-symmetric like linear strategies—they behave differently.
Due to gold’s long-term upward trend in higher timeframes, buy models saw more positive examples during training, which is why they usually perform better than sell models.
Some models only work in ONLY CLOSE mode for sell trades—they do not open new positions, only close existing ones.
Some models trained well for sell-side actions but are not officially recommended yet, as gold’s long-term bullish trend has not reversed.

> This refers to long-term global growth in large timeframes, not short-term fluctuations. Over the past 20 years, gold has grown 1,000%, with only 3 years showing strong downward trends.

Stop-loss is an integral part of the strategy. Many stop-loss activations are present in signals because the EA acts based on the current moment and does not hold losing trades.

This is not a grid system giving $1 per day; profits may accumulate rapidly on one day, or there may be several consecutive stop-loss periods.
Users can join the project during unfavorable phases. Evaluation should span long periods, not just 1–2 weeks.
Models aim to maximize profit per trade, so there are no trailing stops or quick take-profits. Even if a trade could have been closed profitably, intervention is not recommended.

> Every market entry carries risk, regardless of strategy accuracy. Early profit-taking skews the risk/reward ratio unfavorably. Small, frequent profits can create a false sense of reliability, while one or two stop-loss trades can wipe out all accumulated gains. Patience is key.

Gold is an expensive asset, and even 0.01 lots require a minimum deposit. To trade safely with small balances, consider using sub-accounts—this EA supports them.

Project Transparency

All signals are published without prior amplification.
Deposits, withdrawals, or hidden equity lines are not manipulated.
Training-period information is publicly posted in early comments.
When adding new models, any overfitted model is publicly announced and hidden.
Some new models have seasonal annual clustering, but only verified models remain in the main ensemble.
No fake backtests or reviews—only real feedback, signals, and results.

> This project does not promise profits—it is an effort to create a stable tool.