Use ONNX as a trade filter, not a fortune teller.
The robots that print money for retail traders don't exist. The robots that consistently filter bad setups out of an otherwise-edge strategy — those exist. This silo documents the strategy patterns where a small ONNX classifier earns its keep: regime detection, trend confirmation, probability gating.
Strategies, ranked by signal-to-noise.
Market structure classification with ONNX
A three-class model (bull / bear / range) that gates entries from a rule-based EA. The features matter more than the architecture.
read →AI trend filter for a classical EA
Take a working trend-following bot. Add an ONNX layer that suppresses entries when the model says "this isn't a trend." Walk through the integration.
read →Probability-confidence gating in MQL5
The model gives you a probability. The naive code uses the argmax. The robust code waits for confidence to cross a threshold.
read →CNN-LSTM price forecast end-to-end
The canonical CNN-LSTM architecture from research, ported into a working EA. Training script, export, and the MQL5 integration in one piece.
read →Why your model predicts garbage in Strategy Tester
The normalization parameters from training have to be loaded inside the Tester. Most tutorials skip this. It's why your backtest looks like noise.
read →Regime detection: when to turn the EA off
The single most under-rated use of ML in retail trading: knowing when your strategy shouldn't trade. A binary classifier that earns its keep.
read →ONNX as a filter vs ONNX as the whole EA
Two patterns for shipping ONNX in MT5. The filter pattern wins for most retail use cases; standalone is high-ceiling but riskier. Trade-offs and the hybrid most quants actually use.
read →Test it on someone else's capital first.
The strategy survives backtests? Run it through a funded-account challenge before risking your own money. Some firms allow ONNX EAs explicitly.
Compare prop firm rules →