Articles on machine learning in trading

The third article in this series introduces a trainable feature graph in the Cellular10K architecture: the weights of the feature → feature connections are strengthened online after correct predictions and weakened after errors. The article covers soft initialization, the message-passing step, a local Hebbian learning rule, weight clipping, normalization, and decay. The article demonstrates integration with a cellular automaton and a binary predictor, along with diagnostics metrics and practical activation thresholds for controlling overfitting.

In the second part, the cellular automaton is transferred from a grid to a graph. The features become vertices in a graph with local and long-range small-world connections, and the cells become agents that interact not only with their geometric neighbors but also with their semantic neighbors. The article discusses graph-based feature filtering, the construction of a neighbor graph, updated coherence-based voting, and the Graph Coherence and Graph Health metrics. This reduces the impact of isolated outliers and accelerates the propagation of market regimes while maintaining full compatibility with MQL5.

CTradePreflight checks a trade request before OrderSend(), so an EA finds an off-step volume, a wrong-side stop, missing margin or a closed session instead of getting a server rejection. The class repairs only safe cases, reads the OrderCheck() retcode, and counts every refusal by reason. Unit tests and a random-delay stress EA show that levels placed very close to the price can still fail, and an optional delay buffer stops those rejections at the cost of blocking some valid requests.

This article explains why and how the gradient is used in training a perceptron, based on the least squares method and the chain rule for computing partial derivatives. We will demonstrate how to implement gradient calculation in the C_Neuron class using the MQL5 language and test it using examples with 1 and 2 inputs. You'll learn how to adjust the weights and bias using gradient descent and prepare for forward propagation and backpropagation.

We estimate the probability, at broker time 08:00, that EURUSD's 08:00–16:00 range exceeds 1.25 times the mean of the previous 20 complete, valid target ranges. A two-feature SPSS logistic regression model employs the completed 00:00–08:00 H1 window. Coefficients from 2020–2023 sessions are fixed for the 2024 holdout comparison against constant and 60-outcome rolling frequency baselines, using Brier scores and log losses. MQL5 scripts allow readers to extract broker data and recreate the probabilities in MetaTrader 5.

We describe and implement CVO: infection as candidate generation, coordinate-wise normal perturbation, and a dynamic population. The algorithm is integrated into C_AO and has been tested on a standard benchmark. The analysis identifies a scaling-related cause of stagnation and provides a practical solution — a shift to a relative step based on the range width. The code is ready for use.

This article implements a PSI-based drift watchdog for MQL5 EAs. It compares a feature and the model's predicted probability against a quantile-based reference using a rolling window at a configurable cadence, applies hysteresis to avoid flapping, and raises a webhook on state transitions. You get tested math, an audit CSV and a plotting indicator, plus tuning guidance for thresholds, window size, and cooldown.

The article describes TradeMux as a bridge between a Python pipeline and the MetaTrader 5 terminal for the seamless transfer of trading decisions without duplicating logic. The article examines a four-layer production architecture and a complete Python execution service: connection setup, reading account and position data, signal generation (including CatBoost), pre-trade risk control, kill_switch, and a supervisor. Practical benefits include cross-broker normalization (RoboForex, IC Markets, Alpari, OANDA) and scaling from a single account to multi-account broadcasting without changing the trading logic.

This article implements a multi‑timeframe Granger causality monitor for MetaTrader 5. A Jupyter notebook tests predefined timeframe pairs, selects lags, applies multiple‑testing control, and exports a calibration table. An MQL5 Expert Advisor reloads the setup, recomputes the F‑test live on rolling windows, and marks pairs VALID/INVALID to gate its vote‑based BUY/SELL generator. The workflow shows alignment rules, live/offline parity, and a filtered vs unfiltered comparison on EURUSD.

We will create an indicator written entirely in MQL5 to display the activation function and its derivative on the chart, using ReLU as an example. We will explain how derivatives affect the choice of activation function and what precautions are needed at nondifferentiable points. Readers will be able to interactively visualize the curves and gain a practical foundation for deciding when to use certain activation functions and when not to.

This article builds a native MQL5 LOF engine without DLLs, ALGLIB or ONNX. Bars are mapped to five standardized features; k‑NN reachability and density produce a scale‑free LOF score, and the alarm level is a moving quantile of the reference cloud. The score gates only new entries in an EMA‑cross EA, with an indicator for charting and a script for independent checks on XAUUSD, EURUSD and GBPUSD.

The article builds an MQL5‑native weak‑labeling engine that infers per‑indicator trust from vote agreement, then fuses five noisy signals into P(Buy) using Expectation‑Maximization. We describe the EA architecture, rolling‑window refits, abstain handling, diagnostics, and a majority‑vote baseline used for comparison. Readers can reuse the EM library, swap in their own labeling functions, and obtain a transparent, probability‑based entry filter.

This article demonstrates how to encapsulate a neuron in MQL5 using the C_Neuron class, with weights, a bias, and a parameter specifying the number of inputs. We will take a detailed look at how to compute the loss function based on the least squares method and how to organize training data in arrays. This makes it easy to change the inputs and repeat experiments without modifying the implementation.

The article provides a detailed examination of the theoretical foundations and practical implementation of the hidden Markov model with categorical emissions (Categorical HMM) in the MQL5 language. Specific examples are used to demonstrate the processes of inference, iterative parameter learning, and online filtering, as well as the methodology for selecting the optimal model architecture based on the AIC/BIC information criteria.

This article builds an uncertainty-based routing layer for an EA: an online logistic regression scores P(up), Shannon entropy measures uncertainty, and a streaming P² quantile enforces a fixed review budget. Signals auto-trade, auto-skip or go to a human checkpoint, replaced by proxy modes in the Strategy Tester. It is judged on a held-out window against a trade-everything baseline, matched-frequency and seeded controls, and an exact profit split. No edge is claimed.

The article presents a piecewise workflow that forecasts growth in ICA components as surrogate targets and then reconstructs smoothed returns from these forecasts. We analyze dominant-component effects and apply Shapley Additive Explanations to assess feature importance. The approach is implemented end to end, from data preparation to ONNX export and MQL5 integration, providing a practical pipeline for noise reduction and model interpretability.

The article examines the Archimedes Optimization Algorithm — a metaheuristic in which an agent is represented as a physical object with density, volume, and acceleration, and the search itself is reinterpreted as the tendency of bodies submerged in a fluid to reach equilibrium. The balance between exploration and exploitation is not dictated by an external schedule here, but rather stems from the physics of decaying turbulence. We will implement the algorithm in MQL5, test it on a standard test bench, and analyze where this approach works.

Realized backtests reflect a single ordering of trades, so their maximum drawdown and prop‑firm breach risk are often understated. We resequence trade returns via a stationary bootstrap with a derived block length to estimate the drawdown distribution and a sequence risk ratio, compute path‑dependent breach probabilities for drawdown and daily‑loss limits, and replace WFE with a paired IS–OOS degradation statistic. The tools are demonstrated on a real EURUSD RSI strategy to guide sizing and risk limits.