An end-to-end, multi-asset architecture that evaluates, filters, and executes algorithmic strategies using ML meta-labeling and dynamic market regime detection.
Every trade passes through six deterministic layers before execution
A strict, instrument-agnostic pipeline that evaluates every asset independently before aggregating risk at the portfolio level
A rigid backtesting sandbox that mathematically validates trading strategies before live deployment.
Dim_Strategy_RegistryDynamically classifies the current market state for each asset independently using unsupervised learning.
Fact_Market_Regime SQL tablePromoted strategies run continuously on incoming H1 candle data, generating raw trade signals.
XGBoost classification model that contextualizes and filters raw signals. It does not predict price—it evaluates trade probability in the current regime using a dynamic 1:3 ATR-based target variable.
Two-part risk layer: trade-level ATR stops and portfolio-level correlation guard.
Real-time oversight dashboard for human monitoring and system auditing.
Each strategy must pass Layer 0's rigorous mathematical qualification before live deployment
50/200 EMA crossover confirmed by ADX > 25 trend strength filter. Captures momentum in established trends.
Bollinger Band mean reversion combined with RSI < 30 oversold filter. Profits from range-bound markets.
20-period Donchian channel breakouts. Rides momentum when price breaks through established highs/lows.
Stochastic oscillator oversold bounce strategy. Enters when price reaches statistical extremes in ranging conditions.
How we develop, validate, and deploy trading strategies
Strategies encode strict mathematical rules—not black-box predictions. Every entry and exit condition is explicit, reproducible, and testable. We never optimize for past data; we validate against it.
Machine Learning is used strictly as a meta-labeler. The XGBoost model doesn't predict price—it evaluates whether a given strategy historically succeeds in the current market regime. This prevents overfitting to noise.
All backtests account for market regimes. A trend-following strategy is evaluated separately in trending vs. ranging conditions, ensuring we understand precisely when and why strategies work.
We use chronological 70/30 train-test splits with no data leakage. Models are never validated on data they've seen. Temporal features (session flags, multi-timeframe alignment) add real-world context.
Built with production-grade tools designed for reliability and performance
Core language for all pipeline logic, data processing, and ML training
High-performance data manipulation and numerical computation
Gradient boosting for meta-labeling, K-Means for regime clustering
Containerized relational database with ODBC v17 driver
Container orchestration for database and service infrastructure
Technical analysis library for indicator computation (RSI, EMA, ADX, ATR)
REST API integration for live market data and order execution
Development and deployment on Fedora/Ubuntu environments
Real-time dashboards for telemetry and performance visualization
Connected to institutional-grade data feeds and execution
Continuous refinement across all system layers
Layer 5 monitors live expectancy. If a promoted strategy begins decaying, it is flagged for demotion back to Layer 0 for re-evaluation.
The ML meta-labeler's confidence threshold (currently 0.75) is periodically adjusted based on out-of-sample performance metrics.
Expanding the feature set with additional multi-timeframe indicators, volatility surfaces, and cross-asset correlation features.
Continuously developing and testing new algorithmic strategies in the Layer 0 sandbox before promotion to live environments.
Instrument-agnostic architecture allows seamless addition of new forex pairs, commodities, and indices to the pipeline.
Moving toward real-time event-driven architecture for faster signal processing and reduced execution slippage.
Current progress and upcoming milestones
4 strategies qualified with full backtesting pipeline. Evaluator handles multi-asset, multi-strategy testing.
K-Means clustering implemented. Regime ingestion pipeline writing to SQL database.
Self-healing Oanda sync, SQL Server schema, 18+ years of historical H1 data ingested.
Successful batch processing of SQL data for signal generation from promoted strategies.
Optuna hyperparameter tuning for XGBoost, LightGBM, Random Forest, and PyTorch LSTMs.
Implement ATR-based dynamic stops and 30-day rolling correlation matrix.
Streamlit/Power BI dashboard with live regimes, veto rates, and expectancy tracking.
Dynamic ATR Exit Optimization & AI Gatekeeper Tuning (March 8-10)
Live Data Ingestion Automation via CRON (March 12-15)
Vercel Frontend / API Integration (March 18-20)
Paper Trading Sandbox Deployment (End of March 2026)