Articles on machine learning in trading

The article presents a new metaheuristic optimization Circle Search Algorithm (CSA) based on the geometric properties of a circle. The algorithm uses the principle of moving points along tangents to find the optimal solution, combining the phases of global exploration and local exploitation.

We continue to implement the approaches proposed by the authors of the FinCon framework. FinCon is a multi-agent system based on Large Language Models (LLMs). Today, we will implement the necessary modules and conduct comprehensive testing of the model on real historical data.

We invite you to explore the FinCon framework, which is a a Large Language Model (LLM)-based multi-agent system. The framework uses conceptual verbal reinforcement to improve decision making and risk management, enabling effective performance on a variety of financial tasks.

Discover how to fix a critical flaw in financial machine learning that causes overfit models and poor live performance—label concurrency. When using the triple-barrier method, your training labels overlap in time, violating the core IID assumption of most ML algorithms. This article provides a hands-on solution through sample weighting. You will learn how to quantify temporal overlap between trading signals, calculate sample weights that reflect each observation's unique information, and implement these weights in scikit-learn to build more robust classifiers. Learning these essential techniques will make your trading models more robust, reliable and profitable.

In this discussion, we contrast the classical approach to time series cross-validation with modern alternatives that challenge its core assumptions. We expose key blind spots in the traditional method—especially its failure to account for evolving market conditions. To address these gaps, we introduce Effective Memory Cross-Validation (EMCV), a domain-aware approach that questions the long-held belief that more historical data always improves performance.

We continue to develop the algorithms for FinAgent, a multimodal financial trading agent designed to analyze multimodal market dynamics data and historical trading patterns.

We invite you to explore FinAgent, a multimodal financial trading agent framework designed to analyze various types of data reflecting market dynamics and historical trading patterns.

This piece follows up ‘Part-84’, where we introduced the pairing of Stochastic and the Fractal Adaptive Moving Average. We now shift focus to Inference Learning, where we look to see if laggard patterns in the last article could have their fortunes turned around. The Stochastic and FrAMA are a momentum-trend complimentary pairing. For our inference learning, we are revisiting the Beta algorithm of a Variational Auto Encoder. We also, as always, do the implementation of a custom signal class designed for integration with the MQL5 Wizard.

The article introduces the dialectical algorithm (DA), a new global optimization method inspired by the philosophical concept of dialectics. The algorithm exploits a unique division of the population into speculative and practical thinkers. Testing shows impressive performance of up to 98% on low-dimensional problems and overall efficiency of 57.95%. The article explains these metrics and presents a detailed description of the algorithm and the results of experiments on different types of functions.

We will build a biologically correct system of neurons for time series forecasting. The introduction of a plasma-like environment into the neural network architecture creates a kind of "collective intelligence," where each neuron influences the system's operation not only through direct connections, but also through long-range electromagnetic interactions. Let's see how the neural brain modeling system will perform in the market.

In this article, we will forecast future extreme volatility using binary classification. Besides, we will develop an extreme volatility forecast indicator using machine learning.

Layered memory approaches that mimic human cognitive processes enable the processing of complex financial data and adaptation to new signals, thereby improving the effectiveness of investment decisions in dynamic markets.

This is the first part of an article series presenting the implementation of bivariate copulae in MQL5. This article presents code implementing Gaussian and Student's t-copulae. It also delves into the fundamentals of statistical copulae and related topics. The code is based on the Arbitragelab Python package by Hudson and Thames.

This is my own algorithm. The article presents the Time Evolution Travel Algorithm (TETA) inspired by the concept of parallel universes and time streams. The basic idea of the algorithm is that, although time travel in the conventional sense is impossible, we can choose a sequence of events that lead to different realities.

In the previous article, we explored the theoretical foundations and began implementing the approaches of the Multitask-Stockformer framework, which combines the wavelet transform and the Self-Attention multitask model. We continue to implement the algorithms of this framework and evaluate their effectiveness on real historical data.

In the last article, we examined the pairing of Ichimoku and the ADX under an Inference Learning framework. For this piece we revisit, Reinforcement Learning when used with an indicator pairing we considered last in ‘Part 68’. The TRIX and Williams Percent Range. Our algorithm for this review will be the Quantile Regression DQN. As usual, we present this as a custom signal class designed for implementation with the MQL5 Wizard.

In this article, I introduce an innovative trading algorithm that combines evolutionary algorithms with deep reinforcement learning for Forex trading. The algorithm uses the mechanism of extinction of inefficient individuals to optimize the trading strategy.

We invite you to explore a framework that combines wavelet transforms and a multi-task self-attention model, aimed at improving the responsiveness and accuracy of forecasting in volatile market conditions. The wavelet transform allows asset returns to be decomposed into high and low frequencies, carefully capturing long-term market trends and short-term fluctuations.

We have built a robust feature engineering pipeline using proper tick-based bars to eliminate data leakage and solved the critical problem of labeling with meta-labeled triple-barrier signals. This installment covers the advanced labeling technique, trend-scanning, for adaptive horizons. After covering the theory, an example shows how trend-scanning labels can be used with meta-labeling to improve on the classic moving average crossover strategy.

This piece follows up ‘Part-80’, where we examined the pairing of Ichimoku and the ADX under a Reinforcement Learning framework. We now shift focus to Inference Learning. Ichimoku and ADX are complimentary as already covered, however we are going to revisit the conclusions of the last article related to pipeline use. For our inference learning, we are using the Beta algorithm of a Variational Auto Encoder. We also stick with the implementation of a custom signal class designed for integration with the MQL5 Wizard.

We continue our examination of the StockFormer hybrid trading system, which combines predictive coding and reinforcement learning algorithms for financial time series analysis. The system is based on three Transformer branches with a Diversified Multi-Head Attention (DMH-Attn) mechanism that enables the capturing of complex patterns and interdependencies between assets. Previously, we got acquainted with the theoretical aspects of the framework and implemented the DMH-Attn mechanisms. Today, we will talk about the model architecture and training.

We will look at price discretization methods using Python + MQL5. In this article, I will share my practical experience developing a Python library that implements a wide range of approaches to bar formation — from classic Volume and Range bars to more exotic methods like Renko and Kagi. We will consider three-line breakout candles and range bars analyzing their statistics and trying to define how else the prices can be represented discretely.

This article walks the reader through a reimagined version of the classical Bollinger Band breakout strategy. It identifies key weaknesses in the original approach, such as its well-known susceptibility to false breakouts. The article aims to introduce a possible solution: the Double Bollinger Band trading strategy. This relatively lesser known approach supplements the weaknesses of the classical version and offers a more dynamic perspective on financial markets. It helps us overcome the old limitations defined by the original rules, providing traders with a stronger and more adaptive framework.

The article considers a new population optimization algorithm - Cyclic Parthenogenesis Algorithm (CPA), inspired by the unique reproductive strategy of aphids. The algorithm combines two reproduction mechanisms — parthenogenesis and sexual reproduction — and also utilizes the colonial structure of the population with the possibility of migration between colonies. The key features of the algorithm are adaptive switching between different reproductive strategies and a system of information exchange between colonies through the flight mechanism.

The article describes the experience of developing a hybrid trading system that combines classical technical analysis with neural networks. The author provides a detailed analysis of the system architecture from basic pattern analysis and neural network structure to the mechanisms behind trading decisions, and shares real code and practical observations.

This article presents a study of the interaction of different activation functions with optimization algorithms in the context of neural network training. Particular attention is paid to the comparison of the classical ADAM and its population version when working with a wide range of activation functions, including the oscillating ACON and Snake functions. Using a minimalistic MLP (1-1-1) architecture and a single training example, the influence of activation functions on the optimization is isolated from other factors. The article proposes an approach to manage network weights through the boundaries of activation functions and a weight reflection mechanism, which allows avoiding problems with saturation and stagnation in training.

The article describes an innovative approach to forecasting price movements in financial markets using quantum computing. The main focus is on the application of the Quantum Phase Estimation (QPE) algorithm to find prototypes of price patterns allowing traders to significantly speed up the market data analysis.

In this article, we will discuss the hybrid trading system StockFormer, which combines predictive coding and reinforcement learning (RL) algorithms. The framework uses 3 Transformer branches with an integrated Diversified Multi-Head Attention (DMH-Attn) mechanism that improves on the vanilla attention module with a multi-headed Feed-Forward block, allowing it to capture diverse time series patterns across different subspaces.

In the previous article, we introduced the multi-agent adaptive framework MASAAT, which uses an ensemble of agents to perform cross-analysis of multimodal time series at different data scales. Today we will continue implementing the approaches of this framework in MQL5 and bring this work to a logical conclusion.

Dynamic mode decomposition (DMD) is a technique usually applied to high-dimensional datasets. In this article, we demonstrate the application of DMD on univariate time series, showing its ability to characterize a series as well as make forecasts. In doing so, we will investigate MQL5's built-in implementation of dynamic mode decomposition, paying particular attention to the new matrix method, DynamicModeDecomposition().

The article presents the Big Bang - Big Crunch method, which has two key phases: cyclic generation of random points and their compression to the optimal solution. This approach combines exploration and refinement, allowing us to gradually find better solutions and open up new optimization opportunities.

We introduce the Multi-Agent Self-Adaptive Portfolio Optimization Framework (MASAAT), which combines attention mechanisms and time series analysis. MASAAT generates a set of agents that analyze price series and directional changes, enabling the identification of significant fluctuations in asset prices at different levels of detail.