New debiased‑lasso paper lands
A new arXiv paper on Triple/Double‑Debiased Lasso advances high‑dimensional inference for econometric models — a methodological step for consistent estimation and valid confidence intervals in big‑p small‑n settings. This is immediately applicable to factor selection and causal work in quant research. (x.com)
Triple/Double‑Debiased Lasso is authored by Denis Chetverikov (UCLA), Jesper R.-V. Sørensen (University of Copenhagen), and Aleh Tsyvinski (Yale) and was submitted to arXiv on March 20, 2026; the arXiv entry lists the manuscript as 47 pages with 10 figures. (arxiv.org/abs/2603.20134) The paper constructs a Lasso‑based estimator whose moment function satisfies both first‑ and second‑order Neyman orthogonality, explicitly targeting and removing both the leading regularization bias and the second‑order bias. (arxiv.org/abs/2603.20134) The authors derive an asymptotic linear representation for the estimator and show mathematically that the remainder terms for the triple estimator are never larger—and are often of smaller order—than those for the standard double‑Lasso estimator. (arxiv.org/abs/2603.20134) Monte Carlo experiments reported in the paper demonstrate that the triple estimator typically yields more accurate finite‑sample inference and confidence intervals with improved coverage compared with the standard double‑Lasso benchmarks used in the simulations. (arxiv.org/abs/2603.20134) A general recursive formula for constructing higher‑order Neyman‑orthogonal moment functions appears in the paper, framing the triple estimator as a special case and providing a concrete constructive recipe for higher‑order debiasing in Z‑estimation problems. (arxiv.org/abs/2603.20134) The new work builds on the double/debiased machine‑learning and double‑Lasso literature—Denis Chetverikov is a coauthor of the foundational double/debiased ML framework (Chernozhukov et al., 2018)—and follows other recent advances such as the “Doubly Debiased Lasso” approach for hidden confounding (Guo, Čevid & Bühlmann, Annals of Statistics). (arxiv.org/abs/1608.00060 ) (projecteuclid.org) Production ecosystems already exist for double/debiased methods—examples include the Stata ddml implementation and CRAN’s DDL package for doubly debiased Lasso, as well as Python tools like pydoublelasso—potentially easing translation of the paper’s constructive recursive formulas into reproducible code. (statalasso.github.io) (cran.r-project.org) (pypi.org)
