Yevgeniy Koshtenko / Perfil

The article describes TradeMux as a bridge between a Python pipeline and the MetaTrader 5 terminal for the seamless transfer of trading decisions without duplicating logic. The article examines a four-layer production architecture and a complete Python execution service: connection setup, reading account and position data, signal generation (including CatBoost), pre-trade risk control, kill_switch, and a supervisor. Practical benefits include cross-broker normalization (RoboForex, IC Markets, Alpari, OANDA) and scaling from a single account to multi-account broadcasting without changing the trading logic.

The article explores the practical application of the L1 Trend Filter for noise reduction and the construction of structural features suitable for live trading. The full workflow is presented: H1 data for 29 instruments from MetaTrader 5, causal filtering, CatBoost with a horizon of three L1 bars, rigorous walk-forward testing, and VaR-based allocation of lots. The reader is provided with a reproducible code pipeline and a portfolio evaluation methodology.

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.

The article examines the architecture of an Expert Advisor based on a cellular automaton with 10,000 adaptive parameters and an independent binary predictor with a 10-bar prediction horizon. The article demonstrates a three-level online learning process, the evolution of strategies, and validation using a ring buffer and a confusion matrix. The input parameters are reduced to Magic Number, while trading settings are calculated from ATR and five genomes. The EURUSD H1 test yielded an approximate hit rate of 58%, compared to roughly 51% for the fixed MLP.

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.

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 examines the architecture of a committee consisting of 15 AI agents: ten analysts and four risk officers vote in three parallel phases, and the Chair records the final result. Isolated pair contexts with separate reputations are used for eight currency pairs. The dynamic vote threshold depends on the daily PnL targets. The Expert Advisor uses only the signal, SL and TP, which allows for evaluation pf the quality of its decisions without any additional mechanics.

The article discusses the projection of EURUSD daily candlesticks onto the secp256k1 elliptic curve and the extraction of 96 features (EC+TA) to predict the direction of the next candlestick in CatBoost. The price mapping to the curve and the training pipeline on 2,000 D1 bars are shown; the full model achieves a test AUC of 0.6508, with EC features contributing 60.6%. The materials are suitable for reproduction in Python/MetaTrader 5.

An LLM agent with SQL memory in MetaTrader 5 — an architecture that eliminates the main limitation of traditional LLM Expert Advisors (EAs): the loss of context between queries. Each agent decision is recorded in SQLite along with the market conditions at the time of the signal — RSI, MA alignment, volatility — and after the position is closed, the outcome is written to the database. When the next request comes in, the agent reviews its own history — similar past RSI conditions and the last three agent decisions — and only then analyzes the current market. The result is a system that remembers its errors between restarts and adapts its behavior based on accumulated statistics, rather than starting from scratch with each new bar.

The article describes the architecture of a trading system consisting of 20 Grok (xAI)-based LLM agents, each of which embodies a unique trading philosophy — from pure momentum to a statistical z-score. The system applies a genetic algorithm directly during trading: every 20 trades, it automatically eliminates weak agents, clones strong ones with a prompt mutation, and publishes a leaderboard on the MetaTrader 5 chart — without interrupting trading and without using any SDK.

The article describes the transition from four-voice debates to the Council of Fifteen: ten analysts, four independent risk managers, and a Chair, with strict voting rules. It examines the roles of the participants, the three-phase architecture, and the parallel execution of a full cycle in 10–15 seconds. It shows the operational log, the risk-gate rules, and backward compatibility so you can quickly connect the system to an Expert Advisor.

The article describes the architecture of a multi-agent trading system based on the grok-4-fast language model, in which, instead of a single system prompt, four independent analysts with fundamentally different roles operate: a bull, a bear, a risk manager, and an arbiter. Three analysts run in parallel using a ThreadPoolExecutor and, within 3–5 seconds, formulate well-reasoned positions based on the same market data; after that, a deterministic judge renders a final verdict according to strict rules.

The article describes a competitive architecture for MetaTrader 5 in which ten LLM agents, each with different trading rules, manage their own capital and open independent positions using unique magic numbers. The system prompt and the agent's trading aggressiveness are adjusted based on PnL results and the trade streak. A reproducible framework with operating modes and monitored metrics is presented, suitable for testing and further optimization.