Quantitative Trading Platform

Overview

python-trading is a personal research project applying machine learning and quantitative methods to financial markets. It is an operational quantitative trading platform focused on process quality and reproducibility.

Its architecture is organized into dedicated services for data ingestion, signal generation, portfolio optimization, backtesting, execution, and risk controls. The platform uses OpenBB as the provider abstraction layer, computes relative-strength and market-regime signals, maps those signals into Black-Litterman views, and validates portfolio logic through vectorized walk-forward simulation. Operational quality is enforced through automated tests, linting, and type checks, with environment routing that defaults to paper mode before any production promotion. The result is a disciplined engineering workflow for strategy development and execution readiness, rather than opaque discretionary trading.

System Architecture

Machine Learning & Genetic Algorithms

Machine Learning and Genetic Algorithms are complementary AI methodologies used in this platform to solve complex optimization, classification, and forecasting problems in financial markets.

Machine Learning Paradigms

The platform leverages both supervised and unsupervised learning. Supervised learning uses labeled historical data to predict future outcomes — including regression for continuous outputs and classification for discrete signals (e.g., buy vs. not-buy). Unsupervised learning finds hidden structure without labels, commonly used here for clustering assets by risk profile and behavioral similarity.

Model Trade-offs

Interpretability vs. flexibility is a core design trade-off. Simpler models like decision trees are easier to explain, while more flexible models like random forests and neural networks capture richer nonlinear patterns. Boosting families such as XGBoost improve predictive power by combining many weak learners sequentially — each new learner focuses on prior errors, with regularization to reduce overfitting risk.

Feature Engineering

Robust feature engineering is critical. Normalization aligns variables with different scales (price vs. volume) for stable optimization. Detrending removes broad market drift to improve tactical signal detectability. Feature selection uses Boruta-style methods that compare real features against randomized shadow features, keeping all genuinely relevant variables rather than only a minimal subset.

Genetic Algorithms

Genetic Algorithms are evolutionary search methods inspired by natural selection. Candidate solutions are encoded as chromosomes (e.g., entry/exit rule sets). An initial population is scored with a multi-objective fitness function that includes risk, stability, and turnover controls — not just profit. High-fitness candidates reproduce through crossover, random mutation introduces controlled diversity, and over generations robust rule combinations emerge.

Design rules ensure robustness: sufficiently large populations preserve diversity, and iteration limits prevent over-specialization to historical noise.

Practical Integration

Modern workflows combine methods rather than using them in isolation. Unsupervised ML clusters assets by latent structure. Supervised probabilistic ML produces tactical binary states for inclusion decisions. Global optimizers (GA, swarm-intelligence families) tune portfolio-level behavior. LLM layers can enrich numeric models by extracting macro and sector signals from text corpora before those shifts are fully visible in price data.

Signal Frameworks

Governance & Quality

The platform enforces quality through multiple automated gates. Test-driven development with pytest and coverage reporting ensures behavioral correctness. Ruff handles linting and formatting, while mypy provides static type checking across the entire codebase.

Strategy evaluation must report Sharpe, Sortino, and Calmar ratios, drawdown, and turnover impact before promotion to execution phases. Production execution is gated behind stable paper-trading behavior, and walk-forward validation ensures strategies generalize beyond their training period.