Articles on machine learning in trading

Price outliers distort indicators based on the mean and standard deviation. This article delivers a robust MQL5 library (RobustStats.mqh) implementing the median, 1.4826-scaled MAD, and Theil–Sen slope, plus three drop‑in indicators that replace Bollinger Bands, the linear regression channel, and the z‑score oscillator. A comparison overlay and a breakdown‑point measurement on EURUSD show how the robust instruments hold their shape when a single spike moves the classical ones.

We invite you to explore a new approach that combines classical methods and modern neural networks for time series analysis. The article provides a detailed explanation of the architecture and operating principles of the K²VAE model.

We evaluate blind source separation for market noise control using FastICA applied to SMA-filtered, time-lagged OHLC features. The study compares classical and surrogate targets, measures accuracy across lags, tunes KNN models, and inspects residual structure with clustering. Models are exported to ONNX and integrated into an MQL5 Expert Advisor for testing. The result is a reproducible pipeline from data extraction to deployment.

The article discusses the Differential Search Algorithm (DSA), which simulates the migration of a superorganism in search of optimal living conditions. The algorithm uses a Gamma distribution to generate a pseudo-stable random walk and offers four strategies for selecting the direction of movement, along with three coordinate mutation mechanisms. How will this method perform?

We invite you to explore the original implementation of the K²VAE framework — a flexible model capable of linearly approximating complex dynamics in latent space. This article demonstrates how to implement key components in MQL5, including parameterized matrices and how to manage them outside standard neural network layers. This material will be useful for anyone looking for a practical approach to building interpretable time-series models.

We complete the native MQL5 port of Kronos: the decoder, the predictor's decode_s1 and decode_s2 stages with their cross-attention traps, and the autoregressive loop that produces a multi-bar forecast. Then we profile and make it roughly 4.5x faster with an exact KV-cache and pre-transposed weights, verifying every stage against PyTorch.

Trend-scanning supports both forward and backward windows, and the labeling default is unsafe for features: it looks ahead and boosts next-bar agreement well above chance on random walks. We provide a dedicated wrapper, get trend scanning features, that forces computational causal and returns only window, slope, t value, and rsquared. A second analysis quantifies errors introduced by the default log transform on signed series.

We invite you to dive into the exciting world of LightGTS — a lightweight yet powerful framework for time-series forecasting, where adaptive convolution and RoPE encoding are combined with innovative attention mechanisms. In our article, you will find a detailed description of all components — from creating patches to the complex mixture of experts in the decoder — ready for integration into MQL5 projects. Discover how LightGTS takes automated trading to a whole new level!

The article shows how to evaluate machine-learning alphas before a full backtest by expressing them as formulaic alphas. We compute Information Coefficient (IC), Rank IC, Information Ratio (ICIR), and t-statistics to quantify forecasting strength and stability. A MetaTrader 5 backtest illustrates differences versus execution-dependent tests, and a Python parser facilitates reproducible calculations and bulk screening.

This article builds a Kolmogorov–Arnold Network (KAN) in MQL5, where every edge carries a learnable B‑spline curve rather than a scalar weight. We construct the spline basis, assemble edges and a layer, and fit all coefficients by ridge‑regularized least‑squares in a single solve. The model is delivered as an indicator that visualizes the learned curves and an Expert Advisor that acts on the prediction, providing an interpretable, reusable codebase.

We invite you to learn about the innovative technique of adaptive patching — a method for flexibly segmenting time series while taking their internal periodicity into account. We will also look at an efficient encoding technique that preserves important semantic characteristics when working with data at different scales. These methods open up new possibilities for the accurate processing of complex, multiscale data characteristic of financial markets and significantly improve the stability and reliability of forecasts.

Experimental evaluation on standard benchmark functions reveals the advantages and limitations of directly adapting combinatorial algorithms. The article provides a detailed description of the ECEA algorithm's mechanisms and test results.

This article performs a numerical verification of MQL5 eigendecomposition for a covariance matrix using the spectral theorem A = V Λ Vᵀ. It reconstructs the matrix with Diag(), Transpose(), and MatMul(), computes the residual and its Frobenius norm, and shows that deviations remain at floating‑point precision, with results printed to the Experts journal.

The article provides a fascinating look at how SwiGLU embedding reveals hidden market patterns, and how a sparse Mixture of Experts within a Decoder-Only Transformer makes forecasts more accurate at reasonable computational cost. We take an in-depth look at the integration of Time‑MoE into MQL5 and OpenCL, and provide a step-by-step guide to configuring and training the model.

This article builds a white-box classifier in MQL5 using the Tsetlin Machine. It learns human-readable AND-rules instead of weights, trains with integer state updates, and requires no external dependencies. You will assemble the automaton, clause, and multi-class voter, verify on XOR and other boolean tasks, booleanize indicators, label by forward ATR-scaled return, save the model to CSV, and view active rules on a live chart.

We invite you to explore the practical implementation of a sparse mixture of experts block for time series in the OpenCL computing environment. This article provides a step-by-step explanation of how masked multi-window convolution works, as well as how gradient-based training is organized in the presence of multiple information streams.

A new original population-based algorithm, ECEA, is presented. Inspired by the process of water freezing, it adapts ideas from the Crystal Energy Optimizer (CEO) algorithm, which uses graph-based search, for general optimization problems. The algorithm uses a dynamic elite group, three search strategies, and a periodic diversification mechanism.

We invite you to explore the modern Time-MoE framework, which has been adapted for time series forecasting tasks. In this article, we will implement the key components of the architecture step by step, providing explanations and practical examples along the way. This approach will allow you not only to understand how the model works, but also to apply those principles to real-world trading scenarios.

Financial management as an ecosystem: Seven AI traders with different personalities and strategies instead of a single algorithm. They compete for capital, learn from their mistakes, and make decisions collectively. The article explains the principles behind the Modern RL Trader system, in which the code possesses consciousness and emotions, creating a living, evolving trading mind.

The article examines a Williams five‑bar fractal feature pipeline and shows how a centered rolling window creates a true look‑ahead leak. It identifies two additional silent bugs—a hardcoded shift tied to the default n and a volatility threshold that ignores its input—and consolidates fixes under a single leak_safe flag. Readers get leak‑free fractal, level, trend, and signal features, plus guidance on when unshifted columns remain valid for labeling.

This article presents a reproducible MetaTrader 5 workflow: collect history, engineer nine context features, label simulated EMA crossover trades, train with FLAML, and export to ONNX with fixed opset and plain probabilities. The Expert Advisor loads the model natively, mirrors the Python feature contract, and uses a tunable confidence threshold as a trade filter. Readers can swap signals and features to reuse the same pipeline.

The article focuses on the practical implementation of the TimeFound model for time series forecasting. The key stages of implementing the framework's main approaches using MQL5 are examined.

The article examines slice sampling — an adaptive MCMC algorithm that automatically adjusts its sampling parameters. Its effectiveness is demonstrated using Bayesian linear and logistic regression models, and the results are compared with classical frequentist methods.

The article describes the development of a multi-agent machine learning system for algorithmic trading on MetaTrader 5 based on reinforcement learning. The system has a three-tier architecture: memory neurons store experience, agents make independent decisions, and the collective mind combines them through weighted voting. The system is continuously improved through Q-learning, pruning of ineffective neurons, and evolutionary reduction of exploration.

In this article, we build the core of the TimeFound intelligent model step by step, adapting it to real-world time series forecasting tasks. If you are interested in the practical implementation of neural network patching algorithms in MQL5, you have come to the right place.

This article implements BOSS from scratch in MQL5 and applies it to regime classification: SFA turns windows into words, bags record word frequencies, and an ensemble over window lengths votes on labels. We cover the encoding steps, the BOSS distance, training with auto-generated regime labels, and practical parameters. A BTCUSD benchmark versus DTW shows higher macro accuracy on clean data and markedly faster inference.

This article presents two versions of the Crystal Structure Algorithm: the original and the modified version. The Crystal Structure Algorithm (CryStAl), published in 2021 and inspired by the physics of crystal structures, was positioned as a parameter-free metaheuristic for global optimization. However, testing revealed a critical problem with the algorithm. A modified version, CryStAlm, is also presented; it addresses the original's key shortcomings.

The Metropolis-Hastings algorithm is a fundamental Markov chain Monte Carlo (MCMC) method that is widely used to approximate posterior distributions in Bayesian inference. This article describes the theoretical foundations of the algorithm, the implementation of the MHSampler class in MQL5, and examples of its application, including an analysis of the resulting samples.

The Mantis framework transforms complex time series into informative tokens and serves as a reliable foundation for an intelligent trading agent capable of operating in real time.

Kohonen's self-organizing maps transform the chaos of market data into an ordered two-dimensional map, where similar patterns are grouped together. The article demonstrates a complete implementation of a SOM in an MQL5 Expert Advisor with 400 neurons and continuous learning. We break down the Best Matching Unit search algorithm, weight updates using a Gaussian neighborhood function, integration with quantum effects, and the generation of trading signals. The code is open-source, the math is clear, and the results are verifiable.

A metaheuristic algorithm that simulates the growth of coronary arteries in the human heart for optimization problems. It uses the principles of angiogenesis (the growth of new blood vessels), bifurcation (branching), and pruning of weak branches to find optimal solutions in a multidimensional space. Testing its effectiveness across a wide range of tasks yielded unexpected results.

Mantis is a versatile tool for in-depth time series analysis that can be flexibly scaled to accommodate any financial scenario. Learn how a combination of patching, local convolutions, and cross-attention enables a highly accurate interpretation of market patterns.

What if lunar cycles and seasonal patterns influence the foreign exchange markets? This article shows how to translate astrological concepts into the language of mathematics and machine learning. I built a Python system with 88 features based on astronomical cycles, trained CatBoost on 15 years of EURUSD data, and obtained some intriguing results. The code is open-source, the methods are verifiable, and the conclusions are unexpected — ancient wisdom meets gradient boosting.

We present a native MQL5 implementation of the catch22 feature set: all 22 canonical time-series characteristics in a reusable class validated against pycatch22. Using a leak-free pipeline (chronological split, purging, embargo), we run a three-arm ablation—classic indicators, catch22, and combined—for volatility-regime classification. Finally, we deploy the combined model as a Strategy Tester regime filter to quantify its impact on a simple baseline strategy.

Meet Mantis — a lightweight foundation model for time series classification based on a Transformer architecture, featuring contrastive pre-training and hybrid attention that deliver record-breaking accuracy and scalability.