Local Model Inference Runner

Deploying a model trained outside MetaTrader (in Python, R, or anywhere else) usually means either a DLL bridge, which this catalogue's house rules exclude, or a live server round-trip, which adds a network dependency to every tick. This EA instead parses the model's trained parameters directly from a plain text file and evaluates them itself, entirely in MQL4. Two model kinds are supported, chosen by the file's first line. LOGISTIC carries a comma-separated feature name list, a matching comma-separated weight list, and a bias value; inference is the standard sigmoid(bias + weights . features). TREE carries a feature name list followed by one row per node - a node id, a leaf flag, either a feature index and split threshold (internal node) or a predicted value (leaf), and the left/right child ids; inference walks from the root, following the left child whenever the split feature is less than or equal to the threshold and the right child otherwise, until a leaf is reached. Both model kinds read their inputs from a small fixed vocabulary of live-computed features - moving averages, RSI, ATR, momentum, bar shape and volume - looked up BY NAME against the names the model file itself declares, in the order it declares them, so the file controls which of the supported features it actually needs. A model requesting a feature this EA does not know how to compute is reported plainly, not silently substituted with something else. The resulting score drives a conventional entry: at or above InpBuyThreshold with no position already open opens a buy, at or below InpSellThreshold opens a sell, both sized from account equity and risk percent against an ATR-based stop, exactly like the rest of this catalogue's risk-managed Expert Advisors. Notes and limitations The supported feature vocabulary is intentionally small and fixed; a model trained on features outside this list (raw order-book data, cross-asset series, anything this EA cannot compute from the attached chart alone) cannot be deployed here without first retraining against the supported set. There is no validation that a loaded model was actually trained well, or that its training data resembled the instrument/timeframe it is now being run against - a malformed or overfit model will produce a confident-looking score exactly as readily as a well-trained one, and this EA has no way to tell the difference from inside the terminal. Test any new model file with InpDryRun on for a meaningful stretch before trusting it with live orders.

Expert advisor · MT4 · v1.0 · Free

Tags: Buy/Sell Arrows · Risk & Lot Size · RSI · Moving Average · ATR · Momentum · MT4 Expert Advisor · MQL4 Source Code

Expert advisors (EA) for MT4 & MT5 · PolyMaria