The author developed an automated trading system using a modular Go-based monolith. The system retrieves market quotes and order books from MOEX every 5 minutes, calculates technical indicators such as RSI, momentum, and volatility, and checks orders against predefined rules.
The architecture includes a cluster availability feature that switches between a MacBook and an AI-box. The system integrates an ensemble of three models: logistic regression, LightGBM, and XGBoost, which provide a consensus signal. A custom bag-of-words classifier categorizes news into eight specific event types.
Backtesting was conducted on MOEX data from June 19, 2026, to September 16, 2026, using 18 tickers and a deposit of 1 ₽ million. The system includes a kill-switch mechanism that halts operations if the portfolio drawdown exceeds 3%.
Previous experiments involved using LLMs for news sentiment analysis, but they were excluded from the final system due to a lack of determinism and slow backtesting speeds. The author also tested Google's TimesFM model for time-series analysis to identify correlations between different assets.
