Neural Loss-Pattern Auditor - script for MetaTrader 5

Aggregate statistics such as win rate or profit factor cannot show whether the probability of a loss depends on what a trader just did, such as sizing up right after a previous loss. Neural Loss-Pattern Auditor tests for exactly that kind of behavioral pattern, directly on the trader's own trade history.

The main script builds an eight-feature dataset per closed deal and prints a full diagnostic report to the Experts tab:

• Accuracy uplift — the trained network's validation accuracy against the accuracy of always guessing the majority class.
• Calibration table — each probability bin's average prediction against the actual observed loss frequency in that bin.
• Permutation feature importance — which of the eight features the network actually relies on, ranked by the accuracy drop when that column is shuffled.
• A composite score (A to F) combining the three diagnostics, with written recommendations.

Input. With the default InpUseDemoData=true, the script generates 480 synthetic closing deals with a known injected pattern (larger position size right after a loss) and analyzes those, so the output is visible immediately. Set InpUseDemoData=false to read real closed-deal history instead, through HistorySelect() and the HistoryDealGet*() functions; no file or CSV is needed in this mode.

Parameters. InpMinTrades sets the minimum closed trades required before training; InpValidationFraction sets how much of the most recent history is held out for validation; InpHiddenNeurons, InpLearningRate, and InpEpochs control the network; InpCalibrationBins and InpPermutationRepeats control the two diagnostics; and InpWeightUplift, InpWeightCalibration, InpWeightTopFeature plus the three *Scale and four Grade inputs control the composite score.

The tool is written in pure MQL5. It needs no external libraries, no Python, and no AI service of any kind. A full description of the method, the math behind each diagnostic, and the code is available in the article: [ARTICLE_URL].