Sample Splitting and Weak Assumption Inference For Time Series

Abstract

We consider the problem of inference after model selection under weak assumptions in the time series setting. Even when the data are not independent, we show that sample splitting remains asymptotically valid as long as the process satisfies appropriate weak dependence conditions and the functional of interest is suitably well-behaved. In addition, if the inference targets are appropriately defined, we demonstrate that valid statistical inference is possible without assuming stationarity. As a working example, we consider post-selection inference for regression coefficients under a random design assumption, in which the pair $(Y_i, X_i) \in \mathbb{R}^{p_n}$ is assumed to be an observation from a weakly dependent triangular array. We establish (asymptotic) sample splitting validity for regression coefficients under both $\beta$-mixing and $\tau$-dependence assumptions. To facilitate statistical inference in the non-stationary, weakly dependent regime, we extend a central limit theorem of Doukhan and Wintenberger (2007). To extend their result, we derive some properties of the variance of a normalized sum of a weakly dependent process. In particular, we show that, under very general conditions, the variance is often well-approximated by independent blocks. Using this result, we derive the validity of the block multiplier bootstrap under $\theta$-dependence and demonstrate the validity of an inference procedure that combines sample splitting with the bootstrap under weak assumptions.