AI Trading System Development — From Data to a Deployed Model
We build the whole pipeline: data collection, feature engineering, model training, honest out-of-sample validation and deployment into MetaTrader 5 or a Python bot.
An AI trading system uses machine-learning models, such as LSTM or Transformer networks and gradient-boosted trees, to produce trading signals from market data. ExpertNevees builds the full pipeline: data pipeline, features, model training, walk-forward validation, and deployment into MetaTrader 5 through ONNX or into a Python bot. Typical delivery is 21–30 days with 12 months of support. We do not promise profits; we deliver honest, reproducible evaluation.
Model Families We Work With
LSTM / GRU
Transformers
XGBoost / LightGBM
Reinforcement learning
Ensembles
Baselines first
The End-to-End AI Trading Pipeline
Data
Features & Labels
Training & Validation
Deployment
What Responsible AI Trading Development Looks Like
| Topic | Poor practice | Our practice |
|---|---|---|
| Data split | Random train/test split | Chronological splits with walk-forward re-training |
| Leakage | Labels or features that use future bars | Leakage audit of every feature and label |
| Evaluation | Report accuracy only | Report net PnL after costs, drawdown, stability across periods and a comparison to simple baselines |
| Costs | Zero spread and commission | Realistic spread, commission and slippage |
| Claims | Promise returns | No profit guarantee. Reproducible evaluation instead |
| Maintenance | Train once, forget | Drift monitoring and a re-training plan |
Exporting a Trained PyTorch Model to ONNX for MetaTrader 5
import torch
model.eval() # a trained torch.nn.Module
SEQ_LEN, N_FEATURES = 60, 11 # 60 bars of 11 features each
dummy = torch.randn(1, SEQ_LEN, N_FEATURES) # example input that fixes the tensor shape
torch.onnx.export(
model, dummy, "lstm_signal.onnx",
input_names=["candles"], output_names=["probabilities"],
opset_version=17,
)
# The .onnx file can be embedded in an MQL5 Expert Advisor and run with OnnxCreate / OnnxRun.
# Feature scaling used in training must be reproduced exactly inside the EA.MetaTrader 5 can run ONNX models natively, so no Python process is needed on the trading terminal for inference. Feature calculation inside the EA must match the training code bit for bit.
Realistic Expectations About AI in Trading
- ●Financial markets are noisy and non-stationary. A model that worked in the past can stop working, and there is no guarantee of profit.
- ●More complex does not mean better. Simple models with strong validation often outperform complex ones that overfit.
- ●The model is only one part. Risk management, execution quality and costs decide the live result. See Capital & Risk Management.
- ●Plan for re-training. Performance should be monitored live and compared with the validation expectations.
Delivery, Support & Guarantee
Timelines are indicative. The exact price and delivery date are confirmed in writing before work starts, with no hidden costs. Payment: 50% advance, 50% after your approval (USDT or bank wire).
Frequently Asked Questions
Can AI really predict the market?
No model predicts markets reliably. Some models find small statistical edges that can be worth trading after costs, but edges decay. We focus on rigorous validation so you know what the evidence actually supports.
How long does an AI trading system take?
A complete pipeline typically takes 21–30 days. Scope depends on data availability, model type and the deployment target.
Which model is best for trading?
There is no universal winner. We compare baselines, gradient-boosted trees and neural networks on your data and keep the simplest model that performs robustly out of sample.
Can the model run inside MetaTrader 5?
Yes. Models exported to ONNX can run inside an MQL5 Expert Advisor. Alternatively, Python serves predictions to a bot.
Do you provide the training data?
We can source historical data from your broker, MetaTrader or exchange APIs. You need to confirm you have the right to use the data for your purpose.
What do I receive at the end?
Source code, the trained model files, the data pipeline, a validation report with walk-forward results, deployment files and documentation.
Can you review an AI model I already have?
Yes. See Strategy Analysis & Backtest Review for an independent review of leakage, overfitting and cost assumptions.
Official Documentation & Sources
- MQL5 Reference — ONNX ModelsMQL5.comView Docs ↗
- Open Neural Network Exchangeonnx.aiView Docs ↗
Want an AI Trading Model You Can Actually Trust?
Tell us your market, timeframe and data. We start with a feasibility review and honest expectations.