CRYPTO-AI-AGENT: A Personal Financial Analysis Assistant Integrating Machine Learning and Large Language Models

This is an open-source financial analysis system that combines 13 machine learning models, 59 technical indicators, FinBERT sentiment analysis, and a multi-LLM cascading strategy to provide intelligent decision support for the cryptocurrency and stock markets.

• Original Author/Maintainer: Чередняк Юрій Володимирович (Yurii Cheredniak)
• Source Platform: GitHub
• Original Title: CRYPTO-AI-AGENT
• Original Link: https://github.com/yuriieight/CRYPTO-AI-AGENT
• Release Date: May 25, 2026 (v1.0.0 first official version)

CRYPTO-AI-AGENT was initially developed as a graduation project at the Kyiv Polytechnic Institute. Under the guidance of supervisor Мисюк Ірина Володимирівна, author Чередняк Юрій Володимирович combined academic theory with engineering practice to create this open-source tool for individual investors.

The core design philosophy of the project is modularity and extensibility. Traditional financial analysis tools often have single functions—either focusing on technical analysis charts or providing simple price predictions. CRYPTO-AI-AGENT aims to break this fragmentation by integrating data acquisition, feature engineering, model training, prediction generation, sentiment analysis, intelligent Q&A, and portfolio management into a unified architecture. This design approach not only enhances the coherence of the user experience but also reserves sufficient space for future function expansions.

The project adopts a classic front-end and back-end separation architecture. The back-end is built with Python FastAPI to provide high-performance asynchronous APIs, while the front-end uses React + TypeScript + Vite to implement a modern single-page application. The entire system supports one-click deployment via Docker Compose, greatly lowering the deployment barrier.

• Web Framework: FastAPI 0.109 + Uvicorn, providing asynchronous high-performance API services
• ORM Layer: SQLAlchemy 2.0 asynchronous version, supporting PostgreSQL as the main database and SQLite as the research database
• Machine Learning: Mainstream frameworks such as Scikit-learn, XGBoost, LightGBM
• Natural Language Processing: HuggingFace Transformers to load the FinBERT model for financial sentiment analysis
• Large Language Models: Supports multi-level cascading fallback strategy for Claude 3.5, GPT-4o, Gemini, and Llama 3
• Data Caching: Redis 7 for session management and result caching
• Data Sources: CryptoCompare API for real-time cryptocurrency data, yfinance for stock data

• Framework: React 18.2 + TypeScript
• Build Tool: Vite
• Styling: Tailwind CSS 3.4
• Charts: Recharts 2.10 for interactive data visualization

This technology selection balances development efficiency and runtime performance. The asynchronous feature of FastAPI can efficiently handle I/O-intensive tasks (such as calling external APIs, database queries), while React's component-based architecture makes interface development more flexible and controllable.

This is one of the most technically advanced modules of the system. Instead of using a single model, the author built an integrated prediction system containing 13 machine learning models. These models cover different algorithm paradigms:

• Tree Models: Random Forest, XGBoost, LightGBM, CatBoost, Extra Trees
• Linear Models: Ridge, Lasso, Elastic Net
• Support Vector Machines: SVR
• Nearest Neighbor Algorithms: KNeighbors
• Neural Networks: MLPRegressor

During training, the system uses TimeSeriesSplit for time-series cross-validation to ensure the model does not overfit. When making predictions, the system can output results from a single model or a comprehensive prediction from the integrated model. The latter fuses outputs from multiple models via weighted average or voting mechanisms, usually achieving more stable prediction results.

Model evaluation metrics include MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), R² (Coefficient of Determination), and MAPE (Mean Absolute Percentage Error), allowing users to intuitively understand the model's prediction accuracy.

The system has built-in 59 technical analysis indicators, covering multiple dimensions such as trend tracking, momentum analysis, volatility measurement, and volume analysis:

• Trend Indicators: EMA (Exponential Moving Average), SMA (Simple Moving Average), MACD (Moving Average Convergence Divergence)
• Momentum Indicators: RSI (Relative Strength Index), Stochastic Oscillator
• Volatility Indicators: Bollinger Bands, ATR (Average True Range)
• Volume Indicators: OBV (On-Balance Volume), Volume EMA

These indicators are not only used for chart display but, more importantly, as input features for machine learning models. The system automatically calculates these indicator values and incorporates them into the training data of the prediction model to improve prediction accuracy.