Advanced Methods Overview
The sophisticated quant machinery inherited from v1 + de Prado's academic techniques
Horizon’s core stack covers the 90% case. features, signals, Kelly sizing, seven-layer risk, backtest. For the remaining 10% (fund-style research, IC-weighted ensembles, Bayesian hypothesis management, de Prado’s validation methods), there’s a deep layer of advanced machinery inherited from the v1 python/horizon/fund/ modules.
What’s in the box
AlphaModel Multi-factor edge estimation with IC-weighted factor combination. Seven prediction-market factors self-calibrated from rolling information coefficient. SignalEnsemble Combine arbitrary signal sources with dynamic IC-based weighting and redundancy penalties when signals are too correlated. PortfolioOptimizer Constraint-aware Kelly optimizer with Ledoit-Wolf covariance shrinkage and projected gradient descent. HypothesisManager SQLite-backed Bayesian hypothesis lifecycle. prior → evidence → posterior → promote or retire. RiskAnalytics Regime-conditional VaR, portfolio Greeks, stress testing, correlation structure breakdown. PerformanceAttribution Alpha/beta decomposition, factor attribution, per-strategy P&L breakdown. ExecutionIntelligence VPIN toxicity detection, inventory risk tracking, order-flow microstructure analysis. AdaptiveThresholds Self-tuning confidence thresholds from realized outcomes. Strategies calibrate their own conviction levels. PromotionManager Paper → shadow → live promotion gates with automatic demotion on drawdown or degradation. FundCluster Multi-fund orchestration. run many strategies as separate sub-funds with shared risk budgeting. BacktestRunner Fund-aware backtester with synthetic tick generation, walk-forward support, and tearsheet output.
de Prado academic methods
Marcos López de Prado’s Advances in Financial Machine Learning established modern best practices for avoiding overfitting in quant research. Horizon implements or wraps the key techniques:
Validation methods **Purged K-Fold** (prevents label leakage)
**CPCV** (Combinatorially Purged CV)
**Deflated Sharpe Ratio** (multiple-testing adjustment)
**PBO** (Probability of Backtest Overfit) Labeling **Triple-barrier labeling** (profit-take + stop-loss + time)
**Meta-labeling** (learn when to trust a primary model)
**Fixed-horizon labels** with sample weights Portfolio construction **Hierarchical Risk Parity (HRP)**: clustering-based allocation that avoids the inverse-covariance instability of classical mean-variance Features **Fractional differentiation** (preserves memory while achieving stationarity)
**VPIN** (volume-synchronized probability of informed trading)
**Information-driven bars** (tick/volume/dollar bars)