Detecting Sparse Cointegration

1. Problem Statement

The paper addresses the challenge of detecting sparse cointegration in high-dimensional single-equation models.

• Context: Traditional cointegration methods (e.g., Engle-Granger, Johansen) are designed for low-dimensional settings. In high-dimensional environments (where the number of covariates $p$ is large relative to the sample size $n$, or simply large in absolute terms), standard Ordinary Least Squares (OLS) becomes unstable, and identifying the specific subset of variables driving a long-run equilibrium is difficult.
• The Core Issue: Economic and financial data often exhibit "sparsity," meaning a target variable $y_t$ is cointegrated with only a small subset ($s \ll p$) of a large pool of $I(1)$ candidates. However, existing methods struggle to:
  1. Select the correct subset of active covariates when $p$ is large.
  2. Distinguish between a true cointegrating relationship (stationary residuals, $I(0)$) and a spurious regression (non-stationary residuals, $I(1)$) when the residuals are generated from a high-dimensional selection process.
  3. Handle cases where the regressors themselves ($x_t$) may be internally cointegrated, which complicates standard asymptotic distributions.

2. Methodology

The authors propose a two-step procedure that combines modern high-dimensional variable selection with information-theoretic model selection.

Step 1: Sparse Variable Selection via Adaptive Lasso

The first stage estimates the cointegrating vector using the Adaptive Lasso.

• Model: $y_t = \beta_0 + x_t'\beta + z_t$, where $x_t$ is a vector of $p$ $I(1)$ covariates.
• Estimator: The adaptive lasso minimizes the penalized sum of squared residuals:
  $$(\hat{\beta}_0^{AL}, \hat{\beta}^{AL}) = \arg \min_{\beta_0, \beta} \left( \sum_{t=1}^n (y_t - \beta_0 - x_t'\beta)^2 + \lambda_n \sum_{j=1}^p w_j |\beta_j| \right)$$
• Key Features:
  • Adaptive Weights: $w_j = 1/|\tilde{\beta}_j|^\gamma$, where $\tilde{\beta}_j$ is an initial estimator (e.g., OLS). This allows the method to shrink irrelevant coefficients to zero while preserving the "super-consistency" of active coefficients.
  • No Standardization: Unlike standard lasso applications with stationary data, the authors use raw $I(1)$ series without standardization, as the sample mean and variance of $I(1)$ processes are not well-defined in the usual sense.
  • Output: An estimated active set $\hat{S}$ and residuals $\hat{z}_t = y_t - \hat{\beta}_0^{AL} - x_t'\hat{\beta}^{AL}$.

Step 2: Residual-Based Inference via Information Criteria

The second stage determines whether the residuals $\hat{z}_t$ are stationary ($I(0)$) or contain a unit root ($I(1)$).

• Approach: Instead of using traditional unit root tests (like ADF or KPSS) which rely on complex, non-standard asymptotic distributions that depend on the number of regressors, the authors frame this as a model selection problem.
• Competing Models:
  • $M_0$ (Unit Root): $\Delta \hat{z}_t = \mu + \sum \phi_j \Delta \hat{z}_{t-j} + \epsilon_t$ (Restricted, $\phi=0$).
  • $M_1$ (Stationary): $\Delta \hat{z}_t = \mu + \phi \hat{z}_{t-1} + \sum \phi_j \Delta \hat{z}_{t-j} + \epsilon_t$ (Unrestricted, $\phi < 0$).
• Decision Rule: Compare Information Criteria (IC):
  $$IC_{n,k} = \ln(\hat{\sigma}^2_k) + \frac{c_n}{n} \times (\text{number of parameters})$$
  • If $IC_{n,0} > IC_{n,1}$, select $M_1$ (Conclude Cointegration).
  • If $IC_{n,0} \le IC_{n,1}$, select $M_0$ (Conclude Spurious Regression).
• Penalty: The penalty term $c_n$ (e.g., $c_n = \ln n$) is chosen to ensure consistency ($c_n \to \infty, c_n/n \to 0$).

Finite-Sample Adjustment (Capping)

The authors identify a critical finite-sample issue: under spurious regression ($\rho=1$), the adaptive lasso tends to overfit, selecting nearly all $p$ predictors. This causes the residuals $\hat{z}_t$ to appear artificially stationary. To mitigate this, they introduce a parsimony cap ($s_{cap}$) in the implementation, limiting the number of selected coefficients to a user-specified maximum, ensuring the residuals retain their $I(1)$ properties in finite samples.

3. Key Theoretical Contributions

The paper provides rigorous asymptotic theory under a fixed-$p$ framework ($n \to \infty$):

1. Adaptive Lasso Properties (Stage 1):

   • Cointegrated Case ($|\rho|<1$): The adaptive lasso achieves sure-screening (all true active variables are selected) even if the inactive regressors are internally cointegrated. Under stricter conditions (no internal cointegration among regressors), it achieves full model selection consistency.
   • Spurious Case ($\rho=1$): The adaptive lasso fails to select a sparse model; it asymptotically selects all regressors (behaving like OLS) because the spurious correlation prevents coefficients from shrinking to zero.

2. Robustness of Residual Inference (Stage 2):

   • Main Theorem (Proposition 3): The decision rule based on Information Criteria is consistent regardless of the internal cointegration structure of the regressors $x_t$.
   • Mechanism: The validity relies on the fact that under cointegration, $\hat{z}_t$ behaves as $I(0)$, while under spurious regression, $\hat{z}_t$ behaves as $I(1)$. This property holds even if the adaptive lasso selects a non-sparse set of variables in the spurious case.
   • Avoidance of Nuisance Parameters: The method avoids the non-standard limiting distributions typical of residual-based unit root tests, as the decision relies on the relative order of magnitude of residual variances rather than the specific distribution of a test statistic.

4. Empirical Results (Monte Carlo Simulations)

The authors conduct extensive simulations with $p \in \{10, 50, 100\}$ and various cointegration structures (no internal cointegration, internal cointegration among actives, among inactives, and cross-block).

• Stage 1 Performance:

  • The adaptive lasso successfully recovers the true active set in favorable conditions (no internal cointegration).
  • When internal cointegration exists among active variables, exact support recovery drops, but the sure-screening property (retaining all true actives) remains robust.
  • In spurious regression scenarios, the method correctly identifies the need for a large model (over-selection), confirming Proposition 2.

• Stage 2 Performance:

  • Cointegration Detection: The procedure correctly identifies $I(0)$ residuals with near 100% probability when $|\rho| < 1$, even in high dimensions and with complex internal cointegration structures.
  • Spurious Regression Detection: Without the capping mechanism, the procedure frequently fails to detect spurious regression in high dimensions (falsely concluding stationarity due to overfitting). With the capping mechanism, the detection accuracy for spurious regression improves significantly (80–90% accuracy), correctly identifying the unit root.
  • Robustness: The method is robust to endogeneity, serial correlation, and the specific cointegration structure of the regressors.

5. Significance and Conclusion

• Practical Utility: The paper offers a simple, computationally feasible, and robust method for practitioners dealing with high-dimensional economic data (e.g., portfolio selection, factor modeling) where identifying a sparse set of cointegrating variables is crucial.
• Theoretical Advancement: It bridges high-dimensional statistics (lasso) and time series econometrics (cointegration). A key insight is that model selection consistency is not required for valid cointegration testing; the "sure-screening" property combined with the robustness of the residual-based information criterion is sufficient.
• Innovation: By reframing the unit root test as an information-theoretic model selection problem, the authors bypass the need for complex critical values and nuisance-parameter-free distributions, making the procedure more applicable in complex, high-dimensional settings.
• Limitation & Future Work: The current theory relies on fixed-$p$ asymptotics. The authors suggest future work should explore increasing-dimension asymptotics ($p \to \infty$ with $n$) and extensions to multiple cointegrating relationships.

In summary, Gonzalo and Pitarakis provide a two-step, robust framework that effectively detects sparse cointegration in high-dimensional data, overcoming the limitations of traditional tests and standard lasso applications in non-stationary environments.