Convert an XGBoost model to ONNX for MQL5.

XGBoost is one of the strongest baselines for tabular market features, and it deploys to MetaTrader 5 through ONNX just like LightGBM and scikit-learn. Like other tree models, it runs on CPU inside MT5 — no GPU required — which makes it a practical filter to pair with a rules-based EA.

Install

You need xgboost plus the ONNX converters:

pip install xgboost onnxmltools onnxconverter-common onnxruntime

Train a model

import xgboost as xgb
model = xgb.XGBClassifier(n_estimators=200, max_depth=4)
model.fit(X_train, y_train)   # X_train: float32 feature matrix

Convert to ONNX

Use onnxmltools with an explicit float input type. Pin the opset to 17 for MT5:

from onnxmltools.convert import convert_xgboost
from onnxconverter_common.data_types import FloatTensorType
initial_types = [("input", FloatTensorType([None, X_train.shape[1]]))]
onnx_model = convert_xgboost(model, initial_types=initial_types, target_opset=17)
with open("xgb.onnx", "wb") as f:
    f.write(onnx_model.SerializeToString())

Mind the dtypes (the int64 trap)

Feed the model float32 features, not int64. A classifier also emits a label plus a probability tensor — decide which output your EA reads and confirm its name in Netron. This mirrors the scikit-learn int64 trap.

Use it in MQL5 (CPU-only)

Load it like any ONNX model, but keep inference on CPU with ONNX_USE_CPU_ONLY — there is nothing for CUDA to accelerate in a tree ensemble. See the OnnxCreate / OnnxRun reference.

Validate before shipping

Run an ONNX Runtime parity check against the XGBoost prediction on the same input; if they match on the desktop, any MT5 discrepancy is a shape/dtype/normalization issue, not the model.