Walk-forward validation for an ONNX EA.

A single train/test split is the most common reason an ONNX EA looks brilliant in backtest and dies live. Markets change regime; one split just tells you how the model did on one slice of history. Walk-forward validation tests the model the way it will actually be used: train on the past, predict the next unseen window, then roll forward.

Why one split lies

With a single 80/20 split you can tune features and thresholds until the test slice looks great — that is just a slower form of curve-fitting. You also risk lookahead: any normalization or feature computed over the whole dataset leaks future information into the past.

Anchored vs rolling windows

• Rolling: fixed-size train window slides forward (e.g. train 12 months, test the next 1, then shift by 1). Adapts to recent regime; less total data per fit.
• Anchored: train window grows from a fixed start; test the next slice each step. More data, slower to forget old regimes.

The loop

for split in walk_forward_splits(data, train=12*m, test=1*m, step=1*m):
    model.fit(split.train_X, split.train_y)         # fit only on past
    scaler = fit_scaler(split.train_X)              # scaler from train only
    pred = model.predict(scaler.transform(split.test_X))
    record(oos_metric(split.test_y, pred))          # out-of-sample only

Judge the model on the distribution of out-of-sample results across all folds, not the best fold.

Applying it to an ONNX EA

You do not retrain inside MQL5. Walk-forward happens offline in Python; each accepted window produces a new model you export to ONNX, version, and swap into the EA on a schedule. Keep the feature and normalization code identical between Python and the EA — the classic silent break is documented in the normalization bug.

What good looks like

Consistency beats peak. A model that is modestly profitable across most folds is more trustworthy than one that is spectacular in two folds and negative in the rest. Pair it with a confidence filter so the EA simply doesn’t trade when the model is unsure.