Methodology

A five-layer architecture designed to prevent the most common failures in quantitative research: lookahead bias, overfitting, and unrealistic cost assumptions.

05Execution
Cost simulation·Paper verification
04Risk
Position sizing·Circuit breakers·Regime detection
03Signal
ML ensemble·Factor composites·Validation
02Feature
Engineering·Causal windows·Cross-asset
01Data
Point-in-time·Survivorship-free·Gated access
Data flows up · Constraints flow down
01

Data Layer

Point-in-time universe construction
  • Survivorship-bias-free universe: only securities that existed at each historical point are included
  • Chronological gating enforced at the framework level — no future data can leak into past computations
  • Corporate actions adjusted using point-in-time records
  • Multiple data sources with automated quality checks and freshness monitoring
02

Feature Layer

Engineered features, strictly causal
  • Features computed using expanding or rolling windows only — no lookahead by construction
  • Technical, fundamental, and cross-asset feature families
  • Cross-sectional normalization to neutralize outlier effects
  • Automated feature importance monitoring and decay detection
03

Signal Layer

Ensemble models + factor composites
  • Machine learning ensemble methods with strict train/test separation
  • Factor composite signals using cross-sectional percentile ranks
  • Walk-forward validation with multiple-testing corrections on every candidate
  • Overfitting probability estimation before any strategy advances
04

Risk Layer

Position sizing + drawdown control
  • Disciplined position sizing with regime-conditional adjustment
  • Automated circuit breakers at strategy, portfolio, and system levels
  • Risk monitoring with configurable alert thresholds
  • Regime detection drives allocation scaling automatically
05

Execution Layer

Cost simulation + paper verification
  • Mandatory transaction cost model for every backtest — the engine will not run without one
  • Spread, market impact, and commission components calibrated to historical data
  • Paper trading on live market data before any strategy is considered validated
  • Execution quality monitoring: fill rates, slippage, and latency tracking

The architecture is the methodology.

Every bias prevention measure is enforced by the software, not by policy. Chronological gating, mandatory cost models, and out-of-sample validation are structural constraints built into the platform. The system prevents common research failures by design, not by discipline.

Factor Exposure Analysis

Illustrative factor decomposition showing exposure to standard equity risk factors. Not representative of any specific portfolio or time period.

Momentum
+0.65
Market β
+0.45
Quality
+0.32
Value
+0.28
Size
-0.12
Liquidity
-0.15
Volatility
-0.38

Illustrative factor exposures. Not representative of any specific portfolio or time period. See disclosures.