Inference on Time Series Nonparametric Conditional Moment Restrictions Using General Sieves

Xiaohong Chen, Yuan Liao, Weichen WangarXiv 2022 · OpenReview

TL;DR

This paper introduces a new statistical inference framework, the General Nonlinear Sieve Quasi-Likelihood Ratio (GN-QLR), for time series data with conditional moment restrictions. It leverages flexible nonlinear sieve learning (like neural networks) to estimate complex functions, offering a robust way to perform hypothesis testing and construct confidence sets that are asymptotically Chi-square distributed, regardless of the estimability of the functional. This has direct applications in areas like reinforcement learning for off-policy evaluation.

What Problem Does This Paper Address?

In econometrics and statistics, we often encounter models where we have conditional moment restrictions. These restrictions are equations that must hold on average, given certain information. A common challenge is when the functions involved in these restrictions are unknown and potentially very complex (high-dimensional, nonlinear). Traditional methods struggle with such complexity, especially when dealing with time series data where observations are not independent. Furthermore, making valid statistical inferences (like hypothesis tests) in these settings, particularly when the target of interest is a functional of these complex functions, is difficult. The paper addresses the need for a flexible and theoretically sound method to perform inference on these complex, nonparametric functions in time series data, especially when dealing with endogenous variables.

Methods

The core of the paper is the General Nonlinear Sieve Quasi-Likelihood Ratio (GN-QLR) statistic. Here's a breakdown of the key techniques:

  • General Nonlinear Sieves: Instead of traditional linear series expansions, the paper uses "general nonlinear sieve learning spaces." This is a broad category that includes powerful function approximators like multi-layer neural networks and Gaussian radial basis functions. These sieves are more flexible and can better approximate complex, high-dimensional functions, especially those with unbounded supports, which is common in time series.
  • Conditional Moment Restrictions: The framework is built around the idea of conditional moment restrictions, which are expressed as expected values that should be zero.
  • Optimal Weighting: To improve efficiency and robustness, the GN-QLR statistic uses an "optimally weighted" minimum distance approach. This involves estimating a weighting matrix that accounts for the structure of the moment restrictions.
  • Quasi-Likelihood Ratio (QLR) Test: The paper proposes a QLR test for inference. A key theoretical contribution is showing that this statistic is asymptotically Chi-square distributed, a desirable property for hypothesis testing. This holds even for "irregular" functionals that are not root-$n$ estimable, which is a significant advantage.
  • Weak Dependence: The theoretical results are established under a "beta-mixing" condition, which allows for weakly dependent time series data, a common scenario in economics and finance.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and uses synthetic data for simulations.

Key Insights and Ideas

  • Flexibility of Nonlinear Sieves: The paper highlights that using general nonlinear sieve learning spaces (like neural networks) allows for much greater flexibility in approximating complex functions compared to traditional linear methods. This is crucial for real-world data that often exhibits intricate nonlinear relationships.
  • Robust Inference: The GN-QLR statistic's asymptotic Chi-square distribution is a major insight. It means that valid inference (hypothesis testing) can be performed without needing to estimate the Riesz representer (a complex object related to the functional's derivative) or compute standard errors, simplifying practical application.
  • Universality of Asymptotic Distribution: A remarkable finding is that the asymptotic distribution of the GN-QLR statistic does not depend on the specific choice of the nonlinear sieve (e.g., neural network vs. Gaussian radial basis). This means the theoretical guarantees apply broadly, making the method adaptable to various function approximation techniques.
  • Bridging ML and Econometrics: The paper effectively bridges machine learning techniques (neural networks) with rigorous econometric inference, providing a framework for using powerful ML tools in a statistically sound manner for time series analysis.

Key Results

  • The paper establishes the asymptotic Chi-square distribution for the optimally weighted GN-QLR statistic, applicable to both regular and irregular functionals.
  • The theoretical results hold under weak dependence (beta-mixing) conditions for time series data.
  • The method is demonstrated to be applicable to important problems like off-policy evaluation in reinforcement learning, nonparametric instrumental variable regression (NPIV), and nonparametric quantile instrumental variable regression (NPQIV).
  • A Monte Carlo study shows the finite sample performance of the proposed inference procedure.

Limitations and Future Work

  • Tuning Parameters: The authors acknowledge that, as with many machine learning methods, the performance can be sensitive to tuning parameters, especially when applying deep learning for statistical inference. They note that methods using NNs for real applications require more extensive tuning and a good understanding of model complexity.
  • Computational Complexity: While the paper provides theoretical guarantees, the practical implementation and computational cost of training complex neural networks can be a limitation.
  • Future Work: The paper's framework opens doors for further research, particularly in exploring more complex time series structures, developing adaptive methods for parameter tuning, and applying the GN-QLR to a wider range of econometric and statistical problems. The application to reinforcement learning suggests potential for dynamic causal inference.

Summary updated

Abstract

General nonlinear sieve learnings are classes of nonlinear sieves that can approximate nonlinear functions of high dimensional variables much more flexibly than various linear sieves (or series). This paper considers general nonlinear sieve quasi-likelihood ratio (GN-QLR) based inference on expectation functionals of time series data, where the functionals of interest are based on some nonparametric function that satisfy conditional moment restrictions and are learned using multilayer neural networks. While the asymptotic normality of the estimated functionals depends on some unknown Riesz representer of the functional space, we show that the optimally weighted GN-QLR statistic is asymptotically Chi-square distributed, regardless whether the expectation functional is regular (root-$n$ estimable) or not. This holds when the data are weakly dependent beta-mixing condition. We apply our method to the off-policy evaluation in reinforcement learning, by formulating the Bellman equation into the conditional moment restriction framework, so that we can make inference about the state-specific value functional using the proposed GN-QLR method with time series data. In addition, estimating the averaged partial means and averaged partial derivatives of nonparametric instrumental variables and quantile IV models are also presented as leading examples. Finally, a Monte Carlo study shows the finite sample performance of the procedure
BibTeX
@misc{chen2022inference,
  author = {Xiaohong Chen and Yuan Liao and Weichen Wang},
  title = {{Inference on Time Series Nonparametric Conditional Moment Restrictions Using General Sieves}},
  year = {2022},
  eprint = {2301.00092},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.00092},
}
Figure