Daily Summary for 2026-10-20

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: It's the twentieth of October, twenty twenty-six, and this is the day's research.

Jane: 2 new papers came out today.

Tom: I'm Tom.

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

Tom: We'll take the day in one pass, then pull out the papers we're staying with.

Jane: Start us off with Controlling the False Discovery Rate via Model-Agnostic Importance Measures.

The day's research: Tom: Today's focus is improving how we select variables in complex models while keeping track of false discoveries.

Jane: That tracking is vital for making scientific findings interpretable, isn't it?

Lu: Mengtao Wen Haojie Ren Changliang Zou introduces Model-Agnostic Selection with agent-X importance, or MAS.

Meng: MAS uses sample splitting and a new importance measure to bypass assumptions about covariate distributions.

Lalam: It seems to circumvent the need for those assumptions in variable selection.

Tom: This approach aims to provide better false discovery rate control than existing methods.

Jane: How does it evaluate oracle predictiveness across different distributions?

Lu: It evaluates oracle predictiveness across different distributions, which is key.

Tom: Building on that idea, we used folded concave penalties to fix LASSO issues with cointegrated regressors.

Jane: Which specific methods did you employ for the folded concave penalties?

Lu: We utilized SCAD and MCP specifically for this approach.

Meng: How does using those methods resolve the problem with convex LASSO?

Lalam: It achieves a global oracle property, ensuring consistent estimation and better prediction performance.

Lucky paper: 2610.21360: Tom: So, let's look at Controlling the False Discovery Rate via Model-Agnostic Importance Measures. The paper proposes MAS to select variables while controlling the false discovery rate without needing strong assumptions about the data distribution.

Jane: That sounds like it tackles a big problem for interpretable scientific discovery, especially when we use complex machine learning models.

Lu: They introduce Agent-X importance, which uses a resplit-per-covariate strategy and data aggregation to bypass those distributional constraints.

Meng: Bypassing assumptions is interesting, but how much does that change the computational cost compared to methods that rely on known covariate distributions?

Lalam: The paper shows they achieve this by integrating Agent-X importance, a resplit-per-covariate strategy, and data aggregation for cross-fitting.

Tom: Cross-fitting is key there. That sounds like it's designed to decouple the dependence between estimators of importance measures to provide valid false discovery rate control.

Jane: So, the theoretical results establish control under mild conditions when p is less than or equal to n r for some constant r.

Lu: Exactly, Theorem one states that if Assumptions one through five hold and p is at most n r, then the MAS procedure has an FDR of alpha + op(one) as the sample size n and number of predictors p go to infinity.

Meng: That's a lot of assumptions on the data structure, though. What about when we are dealing with very high-dimensional sparse problems where structural assumptions might be easier to leverage?

Lalam: The paper concludes that MAS is recommended in high-dimensional sparse problems where leveraging structural assumptions for regression is often easier than modeling the joint distribution.

Tom: And they did some numerical studies. They found that setting X equal to z zero point nine five strikes a favorable balance between power and false discovery rate control under these settings.

Jane: That specific choice of z zero point nine five seems quite concrete for practical application, but what is the computational efficiency like compared to existing methods?

Lu: The empirical power and much faster computation are observed for MAS compared to other methods like BH-LOCO.

Meng: Faster computation is important when we're shipping models to production, but I worry about real-world traffic. Does this work hold up if the underlying structure changes significantly?

Lalam: The paper shows that the agent-X importance measures are more suitable for the multiple testing problem than LOCO measures in general settings.

Lucky paper: 2610.14617: Tom: So, let's get into Cointegrated Regressors: Where Folded Concave Penalties Go Beyond LASSO. The paper tackles the inconsistency LASSO has when dealing with cointegrated variables because it applies individual l1 penalties to each one instead of looking at the whole system.

Jane: That sounds complicated for someone just trying to predict outcomes, so how exactly does that individual penalty structure cause a problem in prediction performance?

Lu: The mechanism is that the standardized LASSO penalizes coefficients on an individual basis, and because the penalty scales with each regressor according to its stochastic order, an operator of root n is disproportionately large compared to an operator of one.

Meng: So, if we have a cointegrated system where variables move together, this individual scaling just amplifies the shrinkage too much? What does that mean for actual deployment?

Lalam: It means the standard l1 error bounds become invalid because of that mismatch between the penalty and the underlying system structure.

Tom: The authors introduce folded concave penalties, like SCAD and MCP, to be bounded and milder than that unbounded l1 penalty, which they claim avoids overly sparse solutions caused by excessive shrinkage.

Jane: If SCAD and MCP are milder, how do they manage to capture the true coefficients when the data is high-dimensional?

Lu: The theoretical contribution is establishing that these folded concave penalties enjoy the oracle property in high-dimensional predictive regression with cointegration when a tuning parameter is chosen correctly.

Meng: That oracle property sounds good on paper, but what about finding those true coefficients globally? They seem to settle for local optimality at first.

Lu: They address that by proposing constrained FCPs, which minimize the penalized loss function subject to an upper bound on the total penalty level.

Tom: That constraint is key because it favors sparse solutions while avoiding overfitting caused by too many small coefficients, and the global minimizer asymptotically coincides with the oracle estimator.

Jane: So, what is that constraint doing mathematically? How does bounding the total FCP level help control that error term?

Lu: The constraint implies that the l1 error of the estimator is controlled by a small subvector, which lets you use the restricted eigenvalue condition to justify a well-defined global optimizer.

Meng: When you run Monte Carlo simulations comparing MCP and SCAD against LASSO, what were the concrete results on prediction accuracy?

Lalam: The simulations show that folded concave penalties substantially improve parameter estimation and out-of-sample prediction relative to LASSO.

Tom: Specifically, when applying this to predicting the euro area unemployment rate, in fourteen out of eighteen scenarios, a constrained FCP method achieved the lowest root mean square error.

Jane: So, for someone just listening who isn't doing complex regression, what does this mean practically? It means we can trust variable selection more when things are highly correlated.

Lu: The paper concludes by recommending the use of folded concave penalties with a constraint as a robust alternative to LASSO in environments with stationary, nonstationary, and cointegrated time series.