Articles on machine learning in trading

We continue our acquaintance with the Mamba4Cast framework. Today, we will delve into the practical implementation of the proposed approaches. Mamba4Cast was designed not for lengthy warm-up on every new time series, but for immediate deployment. Thanks to the concept of Zero-Shot Forecasting, the model can produce high-quality forecasts on real-world data without additional training or hyperparameter tuning.

The Crow Search Algorithm (CSA) is an elegant metaheuristic inspired by crows’ ability to hide food and find other crows' caches, solving optimization problems by balancing following successful solutions with random exploration of the search space. Let's find out how well the algorithm performs.

A new optimization method, the Bison Algorithm (BIA), uses two strategies, inspired by the behavior of bison, for solving continuous problems with a single objective function. The key features of BIA are two fundamental principles borrowed from the behavior of bison: the ability to move dynamically and a defensive strategy.

In this article we make the case for merging Divergence Mapping with a Temporal Fusion Proxy in a Trading Robot. Rather than depending on lagging price confirmations, the Divergence Mapping's thesis is that acting like a structural sensor can help identify hidden momentum shifts from price action and indicator anomalies. To establish how these anomalies are interpreted over time we use a Temporal Fusion Transformer proxy. This network incorporates historical context to weigh developing trends such that merging it with Divergence Mapping should set us up to spot shifts in accumulation and distribution before price breakouts.

The article describes an interesting approach to algorithmic trading based on symbolic mathematical equations instead of traditional machine learning "black boxes". The author demonstrates how to transform opaque neural networks into readable mathematical equations using the SymPy library and polynomial regression, allowing for a full understanding of the logic behind trading decisions. The approach combines the computational power of ML with the transparency of classical methods, giving traders the ability to analyze, adjust, and adapt models in real time.

In this article we make the case for pairing the Discrete Fourier Transform with a Spiking Neural Network in a Trading Robot. The Fourier Transform helps represent data as oscillations instead of its raw values. To govern how we interpret these cycles, we engage a Spiking Neural Network that unlike regular networks, uses time dependent electrical charges to accumulate potential and only "spike" when a target threshold is met. Combining these two engines allows us better control on the timing of discrete market movements, that in theory should give us entry signals with rigorous mathematical confirmation.

The custom modification of the Dingo algorithm presented in the article has raised the bar for finding the best optimization algorithm. Are even better results possible?

In Part 2 we measure the market's dominant cycle using Ehlers' Hilbert-transform homodyne discriminator and wrap it as an indicator. We then build the MESA Adaptive Moving Average (MAMA) and its follower FAMA from that phase information. Finally, we combine MAMA/FAMA with the Even Better Sinewave to form a regime-switching Expert Advisor and test it on EURUSD in the Strategy Tester, giving you a complete, reproducible MQL5 implementation.

The article presents a new metaheuristic method based on the hunting strategies of Australian dingoes: group attack, chase, and scavenging. Let's see how the Dingo Optimization Algorithm (DOA) performs algorithmically.

We test AFML's claim that the sequential bootstrap decorrelates bagged trees on overlapping triple‑barrier labels by isolating two levers: draw count and draw rule. One decision identical tree is bagged under four row‑sampling regimes and evaluated on EURUSD 2022–2023 for draw uniqueness, between‑tree correlation, AUC, and calibration. Decorrelation comes almost entirely from throttling max_samples to average uniqueness; the sequential draw adds little. Out-of-bag inflation is largest under full-count sequential sampling.

In this article, triangular arbitrage is presented as a problem of finding cycles in a directed graph, where the vertices are currencies and the edges are currency pairs with weight rates. Profitable cycle: product of weights >1. Our Floyd-Warshall and DFS algorithms find optimal currency exchange paths that return to the starting point with a profit.

The article presents a complete implementation of TimeGPT, a specialized Transformer-based architecture for forecasting financial time series on the MetaTrader 5 platform. Adaptation of the attention mechanism to financial data, selective tokenization of price changes, hardware-aware optimizations, and advanced learning techniques are discussed. Included are practical testing results showing 87% forecast accuracy over a 24-bar horizon with a training time of 15 minutes on the CPU. We also present a ready-made trading EA with automatic retraining.

The article presents the revolutionary architecture of PatchTST, a tailored transformer for financial time series analysis that breaks market data into 16-bar patches for efficient processing. We will discuss the full implementation of a trading robot in MQL5 covering everything from mathematical fundamentals and data structures to a ready-made EA with risk management and continuous learning systems.

In Part 2, we introduce a reusable CCovarianceMatrix class that computes and stores a covariance matrix from raw return series using MQL5's native Cov() method. We verify symmetry, print a labeled matrix grid, and call Eig() to obtain eigenvalues and eigenvectors. Readers see how symbols co-move and which factors drive variance, enabling clearer portfolio diagnostics and reuse in scripts or EAs.

In this article, we introduce the Mamba4Cast framework and take a closer look at one of its key components: timestamp-based positional encoding. The article shows shows how time embedding is formed taking into account the calendar structure of the data.

The article discusses the adaptation and practical implementation of the ACEFormer framework using MQL5 in the context of algorithmic trading. It presents key architectural decisions, training features, and model testing results on real data.

In today's article, we will see how a simple code change that makes a neuron slightly more specialized can significantly speed up the training stage. After all, once a neuron or neural network, as we will see later, has been trained, the work it performs becomes much faster. We will also discuss a problem that exists but is rarely mentioned.

We port AFML Chapter 17 structural break tests to MQL5 as a single include, CStructuralBreaks, delivering six bar-indexed features for EAs: CSW statistic and critical value, Chow-Type DFC, SADF with a rolling lookback (default 252), SM-Exp, and SM-Power. SADF uses O(L²) rolling windows for real-time viability. A companion StructuralBreaksViewer indicator plots all series with per‑series visibility and optional z‑score normalization. SB_EMPTY marks invalid values for safe integration.

The article presents an innovative hybrid system for forecasting exchange rates that combines a linear autoregressive model with a U-Transformer architecture for residual analysis. The system automatically switches between signal sources depending on their quality and includes complete trading logic with averaging/pyramiding strategies. The key advantage of this approach is that the neural network is trained on the residuals of the linear model, which simplifies the task and reduces the risk of overfitting. The implementation is done entirely in MQL5 and is ready for use in real trading with automatic adaptation to changing market conditions.

We present a production‑ready implementation of AFML Chapter 17 structural break tests. The module includes Chu-Stinchcombe-White (one-/two-sided), Chow-type DFC, SADF across six models (linear, quadratic, sm poly 1, sm poly 2, sm exp, sm power), plus QADF (q, v) and CADF (q), returning bar-indexed scalar features. We address the book snippets' scaling issues and argument‑order pitfall, and show how a fixed lookback (L=504) bounds SADF cost to O(L²) per bar for regime detection.

We invite you to explore the ACEFormer architecture — a modern solution that combines the effectiveness of probabilistic attention with adaptive time series decomposition. This article will be useful for those seeking a balance between computational performance and forecast accuracy in financial markets.

CTrailingSlidingMedianBiLSTM is a custom MQL5 Wizard trailing module that combines robust median/MAD outlier filtering with a BiLSTM context score in the range [-1, 1]. Four algorithm modes (standard, bands, RSI, adaptive) target noise, mean-reverting bursts and liquidity spikes, reducing premature stop adjustments. This module is intended for side-by-side evaluation with diverse entry signals and money management settings.

What if your trading strategies could learn from each other, like real fighters? Duelist Algorithm is a new optimization method where trading system parameters literally duel for the right to be called the best.

A revolutionary approach to machine learning in trading through quantum computing. The article demonstrates a practical implementation of an adaptive QRC system with continuous retraining for predicting market movements in real time.

We implement Ehlers-style DSP filters in a single reusable MQL5 library and use it to build two indicators. The Roofing Filter applies a 2‑pole high‑pass followed by a Super Smoother to isolate the tradeable 10–48‑bar band. The Even Better Sinewave normalizes the wave to about ±1, oscillating in cycle regimes and railing in trends, so you can read cycles and detect regime shifts in charts and EAs.

The article explores the revolutionary Mamba/SSM neural network architecture for financial time series forecasting. We will consider a complete MQL5 implementation of a modern alternative to Transformer with linear complexity O(N) instead of quadratic O(N²). Selective State Space Models, hardware-aware optimizations, patching techniques, and advanced AdamW training methods are covered in detail. Practical test results showing an increase in accuracy from 62% to 71% while reducing training time from 45 to 8 minutes are included. A ready-made trading EA with auto learning and adaptive risk management for MetaTrader 5 is presented.

The article establishes a reproducible method to measure portfolio risk for multiple symbols using MQL5 matrices and OpenBLAS. It covers computing log returns, building a covariance matrix, and evaluating wᵀΣw instead of summing individual variances. A complete script prints naive versus true volatility and the cross‑term contribution, enabling you to detect when correlated instruments inflate exposure beyond single‑asset estimates.

An MQL5 port of four entropy estimators — Shannon, Plug-In, Lempel-Ziv, and Kontoyiannis — operating on the intrabar tick-rule sequence. CopyTicksRange() limits data to the broker's cached tick window, so features apply to recent bars only. The implementation encodes bid-direction ticks from MqlTick, replaces NumPy-dependent steps with array-based methods, and ships CEntropyFeatures.mqh and EntropyViewer.mq5 for EA and indicator use.

Trading is characterized by high demands on risk management discipline. The article presents an analysis of the main reasons for traders' failures and proposes a technical solution in the form of the CEnhancedRiskManager class for the MQL5 platform. It includes practical testing on an aggressive grid EA.

This article lays out 'CMoneyKDTreeESN' custom money management class usable with the MQL5 Wizard, that combines the KD-Tree algorithm and the Echo State Network. We use the KD-Tree on log returns and ATR to give us a risk score, while the ESN tracks recent flow to give us a bounded lot size multiplier. Our class is usable in a variety of Wizard assembled Expert Advisors as shown here with the Envelopes and RSI signals, with a broad objective of modulating exposure in high-volatility and tail-risk environments.

The article diagnoses two defects that neutralize sequential bootstrap during cross‑validation: type erasure of SequentiallyBootstrappedBaggingClassifier and a fold‑level shape mismatch from cloning full samples info sets. It retains the classifier's identity, adds find seq bagging to re‑inject fold‑sliced t1 in CalibratorCV.fit, and resets state per split. A new bootstrap_comparison module reports OOF and OOB metrics and memory, letting you verify that sequential sampling is applied correctly and quantify its impact.

The article provides production-ready entropy estimators (Shannon, plug-in, Lempel–Ziv, Kontoyiannis) operating on tick-rule–encoded sequences. It resolves three correctness and performance issues in the original code, verifies outputs against chapter references, and extends encoding with quantile and sigma options. Users gain reproducible results and markedly improved computation speed for large bar sets.

In this section, we will look at the implementation of the key interfaces of the library of Gaussian processes in MQL5: IKernel, ILikelihood, and IInference. We will also demonstrate its operation on synthetic data and implement indicators for classification and regression, demonstrating its operation in online mode - with retraining of the model on each new bar.