Articles on machine learning in trading

Deterministic Oscillatory Search (DOS) algorithm is an innovative global optimization method that combines the advantages of gradient and swarm algorithms without the use of random numbers. The fitness oscillation and slope mechanism allows DOS to explore complex search spaces in a deterministic manner.

We build a production MQL5 bet‑sizing toolkit: utilities, snippets, and user‑level functions that mirror the Python originals. The methods cover probability‑to‑size mapping with overlap correction, dynamic forecast‑price sizing (calibrated sigmoid/power with limit price), occupancy‑based budgeting, and mixture‑model reserve sizing (EF3M). The result is a signed [−1, ..., 1] position plus diagnostics you can plug directly into order logic.

We are going to create a matrix forecasting model based on a Markov chain. What are Markov chains, and how can we use a Markov chain for Forex trading?

The Camel Algorithm, developed in 2016, simulates the behavior of camels in the desert to solve optimization problems, taking into account temperature, supply, and endurance. This article also presents a modified version of the algorithm (CAm) with key improvements: the use of a Gaussian distribution in generating solutions and the optimization of the oasis effect parameters.

Integer differentiation forces a binary choice between stationarity and memory: returns (d=1) are stationary but discard all price-level information; raw prices (d=0) preserve memory but violate ML stationarity assumptions. We implement the fixed-width fractional differentiation (FFD) method from AFML Chapter 5, covering get_weights_ffd (iterative recurrence with threshold cutoff), frac_diff_ffd (bounded dot product per bar), and fracdiff_optimal (binary search for minimum stationary d*).

The article contains a detailed description of the cross-rate calculation algorithm, a visualization of the imbalance matrix, and recommendations for optimally setting the MinDiscrepancy and MaxRisk parameters for efficient trading. The system automatically calculates the "fair value" of each currency pair using cross rates, generating buy signals in case of negative deviations and sell signals in case of positive ones.

The bet-sizing signal from Part 10 is concurrency-corrected but carries no payoff-ratio adjustment, no response to a hard drawdown budget, and no validation across combinatorial paths. This article covers three additions: a two-stage architecture in which a Kelly payoff multiplier is applied on top of get_signal, preserving the concurrency correction while incorporating win/loss asymmetry; a prop firm integration layer that calibrates the sigmoid w parameter continuously from the remaining drawdown budget under FundedNext Stellar 2-Step rules; and a CPCV backtest framework that simulates a fresh account state across all φ[N, k] paths, producing a Sharpe distribution and a PBO audit.

Fixed fractions and raw probabilities misallocate risk under overlapping labels and induce overtrading. This article delivers four AFML-compliant sizers: probability-based (z-score → CDF, active-bet averaging, discretization), forecast-price (sigmoid/power with w calibration and limit price), budget-constrained (direction-only), and reserve (mixture-CDF via EF3M). You get a signed, bounded position series with documented conditions of use.

This article introduces Jardine's Gate, a six-gate orthogonal signal filter for MetaTrader 5 that validates LSTM predictions across entropy, expert interference, confidence, regime-adjusted probability, trend direction, and consecutive-loss kill switch dimensions. Out of 43,200 raw signals per month, only 127 pass all six gates. Readers get the complete QuantumEdgeFilter MQL5 class, threshold calibration logic, and gate performance analytics.

The article explores the Battle Royale Optimizer algorithm — a metaheuristic in which solutions compete with their nearest neighbors, accumulate “damage,” are replaced when a threshold is exceeded, and periodically shrink the search space around the current best solution. It presents both pseudocode and an MQL5 implementation of the CAOBRO class, including neighbor search, movement toward the best solution, and an adaptive delta interval. Test results on the Hilly, Forest, and Megacity functions highlight the strengths and limitations of the approach. The reader is provided with a ready-to-use foundation for experimentation and tuning key parameters such as popSize and maxDamage.

This article lays the system architecture for a multi‑account algorithmic trading setup that operates cryptocurrency CFDs on MetaTrader 5 while respecting prop‑firm constraints. It defines three core principles—fixed dollar risk, one script per account, and centralized configuration—then details the Python–MQL5 split, the 60‑second processing loop, and JSON-based signaling. Readers get practical lot‑size computation, safety checks, and position management patterns for reliable deployment.

This article shows how to represent market structure as a graph in MQL5, turning swing highs/lows into nodes with features and linking them by edges. It trains a Graph Neural Network to score potential liquidity zones, exports the model to ONNX, and runs real-time inference in an Expert Advisor. Readers learn how to build the data pipeline, integrate the model, visualize zones on the chart, and use the signals for rule-based execution.

GridSearchCV and RandomizedSearchCV share a fundamental limitation in financial ML: each trial is independent, so search quality does not improve with additional compute. This article integrates Optuna — using the Tree-structured Parzen Estimator — with PurgedKFold cross-validation, HyperbandPruner early stopping, and a dual-weight convention that separates training weights from evaluation weights. The result is a five-component system: an objective function with fold-level pruning, a suggestion layer that optimizes the weighting scheme jointly with model hyperparameters, a financially-calibrated pruner, a resumable SQLite-backed orchestrator, and a converter to scikit-learn cv_results_ format. The article also establishes the boundary — drawn from Timothy Masters — between statistical objectives where directed search is beneficial and financial objectives where it is harmful.

This article implements a box‑constrained Truncated Newton Conjugate‑Gradient (TNC) optimizer in MQL5 and details its core components: scaling, projection to bounds, line search, and Hessian‑vector products via finite differences. It provides an objective wrapper supporting analytic or numerical derivatives and validates the solver on the Rosenbrock benchmark. A logistic regression example shows how to use TNC as a drop‑in alternative to LBFGS.

In the latest installment of this series, we move beyond individual machine learning techniques to address the "Research Chaos" that plagues many quantitative traders. This article focuses on the transition from ad-hoc notebook experiments to a principled, production-grade pipeline that ensures reproducibility, traceability, and efficiency.

What is quantitative trend analysis in the Forex market? We collect statistics on trends, their magnitude and distribution across the EURUSD currency pair. How quantitative trend analysis can help you create a profitable trading expert advisor.