Articles on machine learning in trading

This guide integrates a trained XGBoost model (ONNX) into an SMC EA to evaluate trade setups before execution. The Python pipeline labels historical XAUUSD events and produces a 12-feature representation aligned with the EA. The result is a reproducible method to train, export, and embed the model so the EA can filter OB, FVG, and BOS signals programmatically.

Saplings Sowing and Growing up (SSG) algorithm is inspired by one of the most resilient organisms on the planet demonstrating outstanding capability for survival in a wide variety of conditions.

Category Theory is a diverse and expanding branch of Mathematics which as of yet is relatively uncovered in the MQL5 community. These series of articles look to introduce and examine some of its concepts with the overall goal of establishing an open library that attracts comments and discussion while hopefully furthering the use of this remarkable field in Traders' strategy development.

A large number of the models we have reviewed so far are based on the Transformer architecture. However, they may be inefficient when dealing with long sequences. And in this article, we will get acquainted with an alternative direction of time series forecasting based on state space models.

The article presents a new metaheuristic method based on a fractal approach to partitioning the search space for solving optimization problems. The algorithm sequentially identifies and separates promising areas, creating a self-similar fractal structure that concentrates computing resources on the most promising areas. A unique mutation mechanism aimed at better solutions ensures an optimal balance between exploration and exploitation of the search space, significantly increasing the efficiency of the algorithm.

A multi-task learning framework based on ResNeXt optimizes the analysis of financial data, taking into account its high dimensionality, nonlinearity, and time dependencies. The use of group convolution and specialized heads allows the model to effectively extract key features from the input data.

We continue to explore the innovative Chimera framework – a two-dimensional state-space model that uses neural network technologies to analyze multidimensional time series. This method provides high forecasting accuracy with low computational cost.

We continue studying distributed Q-learning algorithms. In previous articles, we have considered distributed and quantile Q-learning algorithms. In the first algorithm, we trained the probabilities of given ranges of values. In the second algorithm, we trained ranges with a given probability. In both of them, we used a priori knowledge of one distribution and trained another one. In this article, we will consider an algorithm which allows the model to train for both distributions.

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.

In the second part of the article, we will continue developing a modified version of the AOS (Atomic Orbital Search) algorithm focusing on specific operators to improve its efficiency and adaptability. After analyzing the fundamentals and mechanics of the algorithm, we will discuss ideas for improving its performance and the ability to analyze complex solution spaces, proposing new approaches to extend its functionality as an optimization tool.

The Black Hole Algorithm (BHA) uses the principles of black hole gravity to optimize solutions. In this article, we will look at how BHA attracts the best solutions while avoiding local extremes, and why this algorithm has become a powerful tool for solving complex problems. Learn how simple ideas can lead to impressive results in the world of optimization.

This article discusses an approach to trading only in the chosen direction (buy or sell). For this purpose, the technique of causal inference and machine learning are used.

In this article, I will consider the Bat Algorithm (BA), which shows good convergence on smooth functions.

In this article, we will look at the binary genetic algorithm (BGA), which models the natural processes that occur in the genetic material of living things in nature.

What is angular analysis of financial markets? How to use price action angles and machine learning to make accurate forecasts with 67% accuracy? How to combine a regression and classification model with angular features and obtain a working algorithm? What does Gann have to do with it? Why are price movement angles a good indicator for machine learning?

We continue our acquaintance with the TEMPO method. In this article we will evaluate the actual effectiveness of the proposed approaches on real historical data.

In the ever-changing world of trading, adapting to market shifts is not just a choice—it's a necessity. New patterns and trends emerge everyday, making it harder even the most advanced machine learning models to stay effective in the face of evolving conditions. In this article, we’ll explore how to keep your models relevant and responsive to new market data by automatically retraining.

The EURUSD forecasting system with the use of computer vision and deep learning. Learn how convolutional neural networks can recognize complex price patterns in the foreign exchange market and predict exchange rate movements with up to 54% accuracy. The article shares the methodology for creating an algorithm that uses artificial intelligence technologies for visual analysis of charts instead of traditional technical indicators. The author demonstrates the process of transforming price data into "images", their processing by a neural network, and a unique opportunity to peer into the "consciousness" of AI through activation maps and attention heatmaps. Practical Python code using the MetaTrader 5 library allows readers to reproduce the system and apply it in their own trading.

The DeMarker Oscillator and the Envelopes' indicator are momentum and support/ resistance tools that can be paired when developing an Expert Advisor. We continue from our last article that introduced these pair of indicators by adding machine learning to the mix. We are using a recurrent neural network that uses the white-noise kernel to process vectorized signals from these two indicators. This is done in a custom signal class file that works with the MQL5 wizard to assemble an Expert Advisor.

I invite you to explore the MacroHFT framework, which applies context-aware reinforcement learning and memory to improve high-frequency cryptocurrency trading decisions using macroeconomic data and adaptive agents.

Clustering algorithms in machine learning are important unsupervised learning algorithms that can divide the original data into groups with similar observations. By using these groups, you can analyze the market for a specific cluster, search for the most stable clusters using new data, and make causal inferences. The article proposes an original method for time series clustering in Python.

The AI breakthroughs dominating headlines, from ChatGPT to self-driving cars, aren’t built from isolated models but through cumulative knowledge transferred from various models or common fields. Now, this same "learn once, apply everywhere" approach can be applied to help us transform our AI models in algorithmic trading. In this article, we are going to learn how we can leverage the information gained across various instruments to help in improving predictions on others using transfer learning.

We continue the discussion about the use of piecewise linear representation of time series, which was started in the previous article. Today we will see how to combine this method with other approaches to time series analysis to improve the price trend prediction quality.

The last two articles were devoted to the Decision Transformer method, which models action sequences in the context of an autoregressive model of desired rewards. In this article, we will look at another optimization algorithm for this method.

N-BEATS is a revolutionary deep learning model designed for time series forecasting. It was released to surpass classical models for time series forecasting such as ARIMA, PROPHET, VAR, etc. In this article, we are going to discuss this model and use it in predicting the stock market.

Since the first articles devoted to reinforcement learning, we have in one way or another touched upon 2 problems: exploring the environment and determining the reward function. Recent articles have been devoted to the problem of exploration in offline learning. In this article, I would like to introduce you to an algorithm whose authors completely eliminated the reward function.

This is a unique optimization algorithm inspired by the evolution of the turtle shell. The TSEA algorithm emulates the gradual formation of keratinized skin areas, which represent optimal solutions to a problem. The best solutions become "harder" and are located closer to the outer surface, while the less successful solutions remain "softer" and are located inside. The algorithm uses clustering of solutions by quality and distance, allowing to preserve less successful options and providing flexibility and adaptability.

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.

This article applies the Optimal Trading Rule from AFML Chapter 13 to set profit targets and stop-losses without in-sample calibration. We model post-entry P&L with a discrete Ornstein–Uhlenbeck process, run a 100,000-path search, and implement Python, multiprocessing, and a Numba @njit parallel kernel (242× faster). The result is an optimal (PT, SL) under three forecast specifications, constrained by the prop-firm daily loss limit.

Understanding agent behavior is important in many different areas, but most methods focus on just one of the tasks (understanding, noise removal, or prediction), which reduces their effectiveness in real-world scenarios. In this article, we will get acquainted with a model that can adapt to solving various problems.