TradingBot: A Modular Cryptocurrency Quantitative Trading System Integrating CatBoost and Large Language Models

TradingBot is an open-source algorithmic trading framework that integrates traditional machine learning (CatBoost) and large language model (LLM) reasoning capabilities. It adopts a modular architecture design and supports live and paper trading on the Kraken exchange. It aims to address the problems in traditional quantitative trading systems where a single model struggles to capture complex market characteristics and the high conversion cost from research to live deployment, providing an extensible solution for researchers and traders.

The cryptocurrency market has high volatility and 24/7 trading characteristics, providing unique scenarios for quantitative trading strategies. However, traditional quantitative systems face two major challenges: a single model is difficult to capture complex nonlinear market features, and the conversion cost from research to live deployment is high. As an open-source solution, TradingBot decouples data engineering, feature construction, model reasoning, and trade execution through a modular architecture, providing an extensible algorithmic trading framework.

TradingBot uses a layered architecture, divided into data collection layer, feature engineering layer, model reasoning layer, and trade execution layer. Each component can be independently developed and replaced. The data collection layer obtains multi-time granularity historical and real-time data from Kraken and stores it in a standardized way; the feature engineering layer has built-in traditional indicators such as moving averages and RSI, supports custom features, and handles the look-ahead bias problem in financial time series.

The model reasoning layer adopts an ensemble learning approach, combining CatBoost and LLM. CatBoost processes structured market data, learns the relationship between prices and technical indicators, and is efficient in handling categorical features; LLM analyzes unstructured information such as news and social media to extract sentiment signals, which complements technical indicators; the integration strategy uses a dynamic weight mechanism to adjust the weights of the two according to the market environment (higher weight for CatBoost in trending markets, increased weight for LLM in volatile sentiment markets).

The backtesting engine considers friction factors such as transaction costs, slippage, and market impact, supports strategy validation on historical data, and evaluates indicators like Sharpe ratio and maximum drawdown. The system supports paper trading (Kraken simulation environment with no real funds) and live trading, allowing seamless switching after the strategy is stable.

The code follows object-oriented design with clear module interfaces, making it easy to replace components (such as data sources, exchange APIs, and model architectures). System behavior is controlled through configuration files (data parameters, feature selection, model hyperparameters, etc.), so strategy adjustments do not require code modification. A comprehensive logging system records the complete link, supports log level configuration and structured output, facilitating troubleshooting and monitoring.

TradingBot is suitable for three types of users: quantitative research learners (understand the process through source code), strategy developers (quickly validate new ideas), and automated traders (deploy backtested strategies). It is recommended to start with paper trading for first-time use and switch to live trading after understanding the system mechanism; since the cryptocurrency market is highly volatile, strict risk control parameters need to be set.

TradingBot demonstrates the idea of combining traditional ML and LLM to build a complete algorithmic trading system, with modular architecture, configuration-driven design, and integration strategy providing references. In the future, it can further integrate multi-modal capabilities such as chart pattern recognition and video analysis to build a more intelligent trading decision system.