Nonlinear cointegration outperforms linear models
Conventional cointegration tests assume constant adjustment speeds and Gaussian residuals. Empirical evidence rejects both. Our methodology builds statistical arbitrage signals from nonlinear error-correction models that adapt when markets fracture.
Mathematical rigor over marketing claims
Reject overfitted backtests
We do not optimise parameters until a Sharpe ratio looks appealing. Every model must survive an untouched out-of-sample period before it earns the right to be discussed. Curve-fitting is not research. It is a confession of intellectual laziness that we refuse to make.
Document every failure
Abandoned specifications teach more than polished winners. We archive failed models, publish the reasons for their rejection, and use these records to refine our priors. A research culture that hides failure is a culture that will eventually hide blow-ups.
From raw price series to validated signals through rigorous nonlinear testing
Our research workflow is designed to prevent the overfitting and false discovery that plague quantitative trading. Each stage enforces a strict separation between hypothesis generation and hypothesis testing, ensuring that only robust, reproducible models advance toward deployment. We begin with classical Johansen cointegration testing, extend into neural and threshold specifications, and conclude with out-of-sample validation that no internal team can override.
Linear versus nonlinear cointegration
| Features | Staenlandeol | Linear approaches |
|---|---|---|
| Adjustment symmetry | ||
| Constant hedge ratios | ||
| Gaussian residual assumption | ||
| Regime-aware detection | ||
| Robust outlier handling | ||
| Dynamic recalibration | ||
| Structural break adaptation | ||
| Asymmetric reversion speeds |
Threshold and regime models
Technical implementation questions
What software is required?
(Step 2)
Which data frequency works best?
Nonlinear cointegration is most reliably detected at daily or higher frequencies where microstructure noise does not dominate. We have successfully applied the framework to intraday, daily, and weekly data.
How are outliers handled?
We replace ordinary least squares with robust M-estimators and trimmed regression techniques. These methods maintain coefficient stability when heavy-tailed innovations or single extreme observations would corrupt classical estimates.
Can models adapt to regime changes?
Markov-switching cointegration and rolling robust estimators enable dynamic hedge ratios that recalibrate as volatility states shift. Models are also subject to automatic deleveraging when diagnostic tests detect structural breaks.