Articles on machine learning in trading

Most EAs trade more than one setup, yet the Strategy Tester judges them all by a single equity curve. CSetupLedger is a small MQL5 class that stores each trade's setup in its magic number and rebuilds the per-setup record from the account's deal history, so a restart never wipes it. It compares a Wilson interval of each setup's win rate with the breakeven rate implied by that setup's own average win and loss, and gives no verdict before 30 closed trades. Each setup ends up trusted, rejected or still undecided. In two backtests the per-setup totals match the tester exactly. Full source and two working EAs are attached.

This article uses practical examples to explain how to calculate and apply gradients of the weights and bias (intercept) in a linear neuron implemented in MQL5, and also discusses a stochastic variant of gradient descent. We analyze stopping criteria, iteration limits, and the effects of training on data subsets. The results are displayed in the MetaTrader 5 terminal, where a simple chart is also plotted. Readers are invited to modify the training set and analyze the model's behavior.

Introducing the TMA framework — an intelligent system capable of forecasting market dynamics with sufficient accuracy. In this article, we have integrated all the components into a single architecture and transformed it into a fully-fledged trading agent that analyzes the market and makes decisions in real time.

The article focuses on label-auditing for meta-labeling, not on a new entry signal. It presents a native MQL5 implementation of confident learning with out-of-fold scoring and calibrated pruning, plus an EA switch that trains the meta-filter on raw or cleaned datasets. A controlled Strategy Tester comparison helps you quantify noise and choose a pruning level that preserves generalization.

The article continues the series on AI Hedge Fund and addresses three limitations of v4: analysts' reputation scores are now stored persistently in SQLite, the EA has been removed from the critical execution path, and signals from the council of fifteen are distributed to multiple brokers via the TradeMux REST API. The logic behind the council and risk management has remained the same: Python retrieves data via the MetaTrader 5 SDK and executes orders directly. The result is resilience to restarts and scalability across multiple terminals.

Build and deploy a Neural Additive Model natively in MQL5: one small network per feature, additive output, and no post-hoc explainability. The article provides a PyTorch trainer, two‑output ONNX export (signal and contributions), an include wrapper using OnnxCreateFromBuffer/OnnxRun, and an EA that prints real per‑feature contributions each bar. It covers identical feature math for training/export, threshold measurement, Strategy Tester results, and edge cases such as NaNs and out‑of‑distribution inputs.

In this article, theory meets practice. We implement the key modules of the TMA framework — MPE and MPA. Here, data takes on meaning, and cross-attention becomes a tool for precise analysis of market dynamics. Minimal unnecessary operations, maximum efficiency — a step toward next-generation intelligent trading.

The TMA framework offers a new perspective on market dynamics, enabling models to capture not only the state of the market but also the passage of time itself. Its ability to identify patterns in a continuous stream of data makes the analysis more in-depth and accurate than with traditional approaches. Recurrent adaptation turns this method into a practical tool for working with real-world quotes.

We take a closer look at the Color Harmony Algorithm (CHA) — a meta-heuristic optimization method based on Munsell's theory of color harmony. We explore the structure of the hue circle, harmony templates, the alternation between the concentration and dispersion phases, and the role of solution memory. From a theoretical framework to a working implementation in MQL5 and unbiased testing on a standard benchmark.

The article describes the two-level MarkovMLPOscillator indicator: a three-state Markov chain uses historical data to construct a transition matrix and generate 15 probabilistic features for each bar, and an MLP is trained on these features to predict price direction at a specified forecast horizon. The article covers feature generation, the validation procedure using a holdout set, and parameter settings. The result is an interpretable oscillator with a color histogram, a smoothed signal, and a display of the current transition matrix.

Introducing the RAFT framework — a powerful tool for analyzing and forecasting financial time series. Its flexible and optimized architecture ensures accurate forecasts, stable performance, and faster data processing. RAFT reduces the risk of errors and makes it easier to create effective trading strategies.

We present an evolutionary algorithm without a selection operator: the best individual becomes the sole mating partner for the entire population, and classical mutation is replaced by multiplicative mutation with a self-adjusting step size. In this article, we examine three key ideas, implement the algorithm in MQL5 using the C_AO framework, and test it on the standard benchmark suite and in an anti-cheat test — where BOA comes very close to the cutoff for the top 45 but does not make the ranking.

This article measures a live-to-train behavioral-cloning loop for an RL trading system across 83 models. Outcome-weighted imitation shifted the policy toward Hold (48% → 82%), with 12 final models not trading at all. Live records were instead used to fit spread and slippage to broker fills, revealing execution costs ~85× higher than in simulation. We describe the loop, the failure mechanism, the calibration, and a protocol to apply.

The article discusses the fundamentals of Bayesian statistics in both discrete and continuous settings. We will progress from classical Bayes’ theorem and simple examples involving coin flips to conjugate distributions and dynamic Bayesian updating, which enables real-time analysis of market quotes. Using the beta-binomial model as an example, a simple change-point detection indicator has been implemented to help identify shifts in market regimes.

In this article, we explore the inner workings of the RAFT framework — one of the most precise and elegant approaches to analyzing dynamic processes. Step by step, we adapt its idea of iterative refinement to financial time series, laying a solid foundation for a future model. Readers can look forward to an engaging exploration of the architecture, where every component has its own purpose and function.

The article describes an architecture in which the MQL5 Expert Advisor (EA) handles only data collection and execution, while the logic is moved to a Python server with three LangChain agents: a signal agent, a news agent, and a risk manager. Agents process the WebSocket request sequentially; if any agent rejects it, "hold" is returned. Decisions and actual PnL are stored in SQLite, forming memory and statistics. The reader will receive an interaction diagram, a command protocol, and a feedback approach.

Metaprompting is an approach in which the LLM itself optimizes trading instructions based on actual P&L and signal quality metrics. The article demonstrates a practical implementation in Python and MQL5: a prompt version registry, an execution agent, an evaluator for directional accuracy and profit factor, and a meta-LLM that generates improvements in a loop. The solution is integrated into the Expert Advisor without interrupting trading.

This article benchmarks one MLP forward pass implemented three ways in MQL5: native code, ALGLIB, and ONNX Runtime. We quantify numerical fidelity versus a float64 ground truth and execution latency (mean/percentiles), then outline practical criteria and a ready‑to‑use harness to help you choose and integrate the right engine in an EA.

BUZOA is a population-based metaheuristic algorithm in which each agent randomly selects one of three hunting tactics at each iteration: a narrow search around its personal best, a classic PSO step toward the flock leader, or full teleportation to a random point in the search space. The article examines the implementation of the algorithm in MQL5, identifies a coefficient sign error in the original formulation, and presents benchmark results on a standard test bench.

Build and deploy a conservative, offline‑trained policy entirely in MQL5. Logged trades become a fixed dataset for a two‑layer Q‑network trained with Conservative Q‑Learning, which pushes down Q‑values for poorly supported actions to limit extrapolation error. The EA loads the frozen weights and applies an out‑of‑distribution guard using saved feature statistics. This yields reproducible training and execution without live exploration.

The article presents a regression-based approach to forecasting Renko bars using CatBoost: the model estimates the logarithmic return of the next bar and the uncertainty of the forecast. We will examine a cascade of residual models with OOF validation using TimeSeriesSplit, shrinkage, and overall early stopping, as well as conditional bias correction. On EURUSD D1, a reduction in OOF-MAE was achieved, along with approximately 65% directional accuracy. A working script for MetaTrader 5 is provided that generates a signal, position size, SL, and TP in brick units.

The article presents the comprehensive architecture of the STE-FlowNet encoder, which combines stack memory, recurrent processing, and a correlation mechanism to extract hidden market dependencies. The article demonstrates how these modules are sequentially integrated into a single computational chain capable of performing comprehensive time-series analysis.

The article constrains the optimal trading rule by a prop-firm stop-loss ceiling and corrects OU fitting by pooling within-trade pairs per AFML eq. 13.6. We apply the ceiling as a feasibility mask on a shared PT–SL mesh, ensuring the constrained set is a subset of the baseline. Practical guidance covers when the ceiling binds, the Sharpe cost, and how exposure and live budget state affect the bound.

The algorithms were rerun on the updated functions, and a method was proposed for quickly verifying their "honesty". The composite test combines five different landscapes and eliminates the possibility of gaining an advantage based on the geometry of individual problems, allowing for a quick assessment of the algorithm's actual search capability. A script is provided for preliminary validation of algorithms before they are applied to the optimization of trading strategies.

We analyze why ratings may be inflated due to the alignment of algorithm trajectories with benchmark diagonals, and we supplement the testing methodology with a requirement to keep the global extremum away from the diagonals. We update Forest and Megacity, and perform RAW verification and calibration using VerifyExtremes.mq5. A drop in HHO and DOAm performance serves as a practical indicator of false leaders.

A complete Hyperdimensional Computing library for MQL5 implements bind, bundle and permute over 4096‑bit hypervectors, graded level ladders for quantities, an encoder for bars and sequences, and a classifier trained by integer addition. It includes a purged, ordered validation split, online updates without refitting, a market scan with two nulls, and a Python reference that reproduces the MQL5 model bit for bit, so you can validate results and deploy a reproducible indicator and test suite.

The article shows how to compare two trading forecasters in MQL5 using the Diebold–Mariano test with HAC‑corrected variance and the Harvey–Leybourne–Newbold adjustment. It explains implementation details, including loss choices, horizon‑dependent lag windows, rolling computation, verdict handling, and a manifest‑driven configuration, plus a Student‑t implementation without external libraries. An EA and an indicator provide a live dashboard and significance bands for practical, defensible model selection.

The STE-FlowNet framework offers a new perspective on financial data analysis by responding to actual market events rather than fixed timeframes. Its architecture preserves local and temporal dependencies, making it possible to track even small impulses in price dynamics.

The article rebuilds a Gaussian Process forecaster in native MQL5 around market-state inputs and direct horizon log-return targets. It employs a Matern 3/2 kernel, symmetric Cholesky with adaptive jitter, standardized targets, and a log‑space grid search that avoids state side effects. The EA trades only when az‑score over observed variance clears a threshold and the edge beats the live spread. Calibration is verified in Strategy Tester.