This paper adapts the Double Machine Learning (DML) estimator for macroeconomic time series, which are typically short, persistent, and highly endogenous. The key innovation is Reverse Cross-Fitting (RCF), a deterministic cross-fitting scheme that leverages the time-reversibility of stationary processes. Unlike random cross-fitting, RCF preserves the temporal structure by training nuisance functions on either past or time-reversed future data, avoiding the sample loss of block-based methods like Neighbors-Left-Out (NLO). The authors prove that under conditions of weak dependence, Neyman orthogonality, and conditional stability, the RCF-DML estimator is root-T consistent and asymptotically normal, with a long-run variance that can be consistently estimated via HAC methods.

A second contribution addresses hyperparameter tuning for nuisance learners. Standard predictive tuning (e.g., minimizing RMSE) can lead to high bias in high-dimensional settings, as it does not align with the bias-variance trade-off required for valid causal inference. The paper proposes a stability-based criterion that identifies a 'Goldilocks zone' of tuning parameters where the fold-specific RMSE varies minimally. Within this zone, the estimator achieves minimal bias, even when predictive RMSE is not minimized. Extensive simulations based on structural VAR and partially linear regression DGPs show that RCF-DML performs well in finite samples, outperforming NLO under strong persistence and remaining robust to model misspecification, heteroskedasticity, and non-normality. An empirical application to the dynamic effects of Tier 1 regulatory capital shocks demonstrates the method's practical utility, yielding results consistent with prior literature.