Online Learning Workflows for Updating Trading Models in MetaTrader 5
Summary
The document explains online learning as a way to update predictive models as new market observations arrive. It contrasts this approach with batch training and discusses adaptability to changing patterns, lower memory demands for incremental methods, and the potential for more timely predictions. A MetaTrader 5 workflow is described: a Python client retrieves recently closed bars, creates a next-bar direction target, trains a classifier, and saves it in ONNX format for use by MQL5 programs.
The article also outlines updating CatBoost and GRU models in chunks, including preprocessing and saving model artifacts to a shared folder. It reports example chunk accuracy scores ranging from 0.455 to 0.565, which are close to chance for a binary direction task and do not establish profitability. The examples are implementation guidance rather than a validated trading strategy. The document gives little detail on safeguards against leakage, drift detection, update scheduling, transaction costs, or how to evaluate the resulting models in a realistic trading simulation.
Key ideas
- Online learning updates a model incrementally as new observations become available.
- A Python client can retrieve recently closed MetaTrader 5 bars and prepare model inputs and targets.
- The described workflow trains classifiers and exports them in ONNX format for MQL5 use.
- Chunk-based updates are illustrated for CatBoost and GRU models, with preprocessing artifacts stored alongside the models.
- Reported classification accuracy varies around chance and does not demonstrate trading performance.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.