Cointegrated Regressors: Where Folded Concave Penalties Go Beyond LASSO

arXiv:2610.14617 · stat.ME, econ.EM, stat.ML · Submitted 2026-10-10 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Cointegrated Regressors: Where Folded Concave Penalties Go Beyond LASSO".

Jane: The gist: Folded concave penalties (FCPs) like SCAD and MCP are proposed as remedies to LASSO's failure in consistently estimating coefficients of cointegrated regressors in high-dimensional predictive regressions.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: To wrap up on "Cointegrated Regressors: Where Folded Concave Penalties Go Beyond LASSO", the core message is clear: LASSO fails to consistently estimate coefficients when cointegration is present because it ignores the shared system structure.

Jane: So, the authors propose using folded concave penalties, like SCAD and MCP, to solve this by giving us a local property and a global property for the estimation.

Lu: The most important result they push forward is that with a penalty-level constraint on these FCPs, you get a global minimizer that consistently estimates the true parameters and achieves oracle prediction performance.

Meng: From an operational standpoint, this means we have a more robust method than LASSO for high-dimensional problems involving time series where long-run relationships are present.

Tom: It moves beyond just variable selection inconsistency and tackles the actual parameter estimation consistency problem that happens when you have these linked predictors.

Jane: This work provides a foundation for using constrained FCPs as a strong tool for out-of-sample prediction in many complex economic forecasting environments.

Conclusion: Tom: So we're wrapping up this session on "Cointegrated Regressors: Where Folded Concave Penalties Go Beyond LASSO." Basically, they took a common method and showed why it breaks when you have these linked time series.

Jane: Right. The paper argues that standard LASSO just gets confused when variables are cointegrated because it treats them all individually instead of looking at the whole system together.

Lu: And they introduce these folded concave penalties, like SCAD and MCP, to fix that by making the penalty behave better for these linked structures.

Meng: It’s about avoiding that excessive shrinkage LASSO does on the variables that are actually tied together in a cointegration relationship.

Tom: So if we distill it down, the main point is they establish a way to estimate those coefficients consistently when you have these tricky time series dependencies.

Jane: They do this by showing two things: first, a local property where their method captures the best possible fit under certain conditions.

Lu: And second, a global property achieved when you add a constraint on the total penalty level of these FCPs.

Tom: That constraint part is what really gets me because it moves it from just being good locally to actually finding the true coefficients globally across the whole model.

Jane: So what does this mean for us? It means we have a more reliable way to predict things when we know there are these long-run relationships between our data points.

Lu: It gives us a solid theoretical foundation for using these constrained methods in real high-dimensional applications where cointegration is present.

Meng: For production, it suggests that if you can manage the computational cost of the constraint, this approach offers better out-of-sample results than just sticking with LASSO on your own.

Tom: Exactly. It’s about moving from a method that has these inconsistencies to one that actually delivers stable estimates under real-world complexity.

Jane: Next time, we'll look at how these specific algorithms actually perform in simulation studies against the standard LASSO benchmark.

Hongqi Chena, Chengwang Liaob, Ziwei Meic, Zhentao Shib

Hunan University · The Chinese University of Hong Kong · University of Macau

stat.ME, econ.EM, stat.ML

Submitted: 2026-10-10

Updated: 2026-10-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist: Folded concave penalties (FCPs) like SCAD and MCP are proposed as remedies to LASSO's failure in consistently estimating coefficients of cointegrated regressors in high-dimensional

Key concepts

LASSO Failure under Cointegration: Standard LASSO penalizes each variable individually based on its own scale. In cointegrated systems, this ignores the relationship between variables, causing excessive shrinkage of the coefficients for those linked by cointegration. This mismatch leads to poor out-of-sample prediction because the penalty is disproportionately large relative to the actual signal.
Folded Concave Penalties (FCPs): FCPs, such as SCAD and MCP, are convex penalties designed to be bounded and much milder than the unbounded L1 penalty used by LASSO. They prevent overly sparse solutions caused by excessive shrinkage. These penalties enable better variable selection while maintaining the desirable oracle property in high-dimensional settings.
Constrained FCP Regularization: To find the true global coefficients, the authors propose minimizing the penalized loss function subject to an upper bound on total FCP. This constraint favors sparse solutions and prevents overfitting from many small coefficients. The resulting global minimizer asymptotically matches the oracle estimator, allowing for consistent coefficient estimation.

Summary

The gist: Folded concave penalties (FCPs) like SCAD and MCP are proposed as remedies to LASSO's failure in consistently estimating coefficients of cointegrated regressors in high-dimensional predictive regressions.

Inconsistency of LASSO under Cointegration

LASSO fails to deliver consistent parameter estimation when facing cointegrated variables because it imposes individual l1-penalties on the coefficients, ignoring the underlying cointegrated system, which causes excessive shrinkage of the coefficients of cointegrated regressors and invalidates standard l1 error bounds leading to deteriorated out-of-sample (OOS) prediction performance. This inconsistency is rooted in the mismatch where the standardized LASSO penalizes coefficients on an individual basis, rescales the penalty according to the stochastic order of each regressor, which results in an Op(√n) penalty is disproportionately large relative to the Op(1) signal. Specifically, a bivariate toy model demonstrates this failure where the LASSO criterion function is shown to be "much larger than L LASSO λ(bϕa = 0, bϕb = 0) = n−1 Pn t=1 y 2t + 0 p→ E(y 2t) < ∞".

The Mechanism of Folded Concave Penalties (FCPs)

FCPs, such as SCAD and MCP, are introduced to address LASSO's shortcomings by employing folded concave penalties (FCP), including the well-known SCAD and MCP. These penalties are designed to be bounded and much milder than the unbounded l1-penalty, which avoids overly sparse solutions caused by excessive shrinkage, enabling oracle variable selection. A key theoretical contribution is establishing that FCPs enjoy the oracle property in the high-dimensional predictive regression with cointegration, where a properly chosen tuning parameter ensures that the oracle least squares estimator is captured, asymptotically, by one of the local optimizers of the regression penalized by FCPs.

Constrained Regularization for Global Optimality

To move beyond local optimality and find the true coefficients, the authors propose tailoring the regularization by minimizing the penalized loss function subject to an upper bound on the total FCP, referred to as constrained FCP. This constraint is crucial because it favors sparse solutions and avoids overfitting caused by many small coefficients, and the final theoretical result shows that the global minimizer of the constrained scheme asymptotically coincides with the oracle estimator. This is achieved because the constraint implies that the l1 error of the estimator is controlled by a small subvector, which allows for the use of the restricted eigenvalue (RE) condition to justify a well-defined global optimizer that consistently estimates the true parameter.

Theoretical Guarantees and Conditions

The theoretical framework relies on several assumptions, including Assumptions 1 through 6 which characterize the stationary innovations and their dependence structure. The paper establishes a set of conditions for the penalty function Jλ(t) under Condition 1, ensuring it satisfies properties like Jλ(t) ≥ a1λt for any t ∈ [−aλ, aλ] and Jλ(t) ≤ C λ2 for any t ∈ R. Furthermore, Assumption 7 imposes a condition on the minimum gap of nonzero coefficients to ensure the nonzero coefficients are sufficiently large and distinguishable from zero, which is necessary for consistent variable selection.

Empirical Validation

Monte Carlo simulations comparing LASSO, MCP, SCAD, and their constrained versions (MCP-C and SCAD-C) show that the FCPs substantially improve parameter estimation, variable selection, and OOS prediction relative to LASSO. In the application to predicting the euro area unemployment rate, the FCPs consistently outperform LASSO, and in 14 out of 18 scenarios, the lowest RMSE is achieved by a constrained FCP method. These simulation results confirm the theoretical predictions regarding the robustness of FCPs in high-dimensional cointegrated settings.

Conclusion

The paper concludes by recommending the use of FCPs with a constraint to bound the total penalty level as a robust alternative to LASSO in environments containing stationary, nonstationary, and cointegrated time series. The theoretical results establish coefficient estimation consistency and the oracle property for the recommended procedure when stationary, nonstationary, and cointegrated time series are present. This work provides a foundation for using constrained FCPs for OOS prediction in high-dimensional cointegrated regressions.

How it works

**(The summary above covers the required structure and content based on the provided text.

Improvements for AI systems

textbfStronger Predictive Consistency for Cointegrated Regressions using Constrained Penalties (MCP-C/SCAD-C): The proposed method guarantees the global minimizer of the constrained scheme asymptotically coincides with the oracle estimator, which enjoys a global solution that consistently estimates the true coefficients and enjoys oracle prediction performance. This means AI systems can be deployed in high-dimensional settings where variables are cointegrated without risking inconsistent parameter estimation, leading to superior out-of-sample prediction compared to LASSO.
textbf Enhanced Variable Selection Accuracy: The results show that FCPs substantially improve variable selection relative to LASSO, with simulations demonstrating variable selection accuracies close to one for the constrained versions. AI systems can therefore achieve a more accurate and reliable identification of the true sparse parameter set when dealing with complex, cointegrated data structures.
textbf Robustness Against Nonstationarity: The theoretical framework addresses mixed data settings (stationary, nonstationary, and cointegrated variables), establishing that FCPs can be a robust alternative to LASSO with regressors of diverse dynamic behaviors. This allows AI models to effectively utilize a vast information set containing unit roots and persistent processes without the penalty structure collapsing due to the mismatch between individual regressor orders.
textbf Adaptive Model Selection via Data-Driven Tuning: The paper recommends selecting tuning parameters data-driven through cross-validation, specifically for SCAD and MCP-C, which minimizes the penalized loss function in (4). This capability allows AI systems to automatically determine the optimal regularization strength based on empirical data characteristics rather than relying on fixed theoretical constants.