ARV ML Dataset Export — User Manual

1. What it does

ARV ML Dataset Export turns price history into a machine-learning dataset that is ready for training an ONNX model outside the terminal. One run on any chart produces two files:

The script reads history and writes files into the terminal sandbox. It never trades, never places orders and never sends data anywhere.

2. Why it is built this way

Two design decisions matter more than the code itself.

Normalization is fitted on the training window only. Fitting mean/std on the whole history and then training on a part of that same history leaks future information into the model: the features of 2016 are scaled by statistics computed from 2026. The exported JSON therefore also contains full_means, full_stds and a causal_check block that quantifies the difference. If the difference exceeds the tolerance you set in InpCausalTolPct, the journal prints a FAIL line and you know that full-sample normalization would bias the model.

The train/test split is a date, not a random shuffle. Financial series are not independent; a random split puts neighbouring bars from the same market regime into both parts. InpSplitDate keeps the split chronological.

3. Installation

1. Copy ARV_ML_Export.ex5 to <terminal data folder>\MQL5\Scripts\.
2. In MetaTrader 5 press F5 in the Navigator so the new script appears.
3. No libraries, no DLL, no WebRequest, no external services are used.

4. Running the script

1. Open a chart of the instrument you want to export. The chart period does not matter: all data is taken from H1, H4 and M15 series of that symbol.
2. Drag ARV_ML_Export from the Navigator onto the chart, or double-click it.
3. Check the inputs and press OK.

Wait for the line ARV ML Export: done, ... rows in the journal (Toolbox → Experts).

5. Inputs

6. Output format

CSV header (comma separated, ASCII):

time,f0,f1,f2,f3,f4,f5,close,atr,dir,label,mfe_r,mae_r,split

Feature definitions (all values as of the closed H1 bar; H4 and M15 values are taken from the last bar closed at that moment):

f0 = |EMA20(H1)  - EMA50(H1)| / ATR14(H1)
f1 = (EMA20(H1)  - EMA20(H1) 5 bars ago) / ATR14(H1)
f2 = (Close(H1)  - EMA200(H1)) / ATR14(H1)
f3 = (BB_upper(H4) - BB_lower(H4)) / EMA20(H4)          Bollinger 20, 2.0
f4 = ATR14(M15) / Close(M15) * 100
f5 = StdDev20(H1, SMA) / EMA20(H1)

JSON keys, in order:

version, means[6], stds[6], schema, symbol, timeframe, split_date,
rows_train, rows_test, features[], full_means[6], full_stds[6],
causal_check{max_rel_diff, threshold, passed}

7. Using the dataset for training

Training is done outside the terminal, for example with Python and NumPy: load the CSV, normalize the features with means/stds from the JSON, train a small network and export it to ONNX. The export is deliberately framework-agnostic: any tool that can read CSV, JSON and write ONNX will do.

An example project that also writes the ONNX file without any ONNX package (a small protobuf writer) is published with the companion article; the models it produces load in MetaTrader 5 through OnnxCreateFromBuffer.

8. Troubleshooting

9. Disclaimer

The script performs no trading and gives no signals. It only produces data files. Backtest results and machine-learning models built on this data do not guarantee future performance.