Abstract:
This talk concerns estimation of and testing for a class of locally stationary time series factor models with evolutionary dynamics, where the entries and dimension of the factor loading matrix are allowed to vary with time while the factors and idiosyncratic noise components are locally stationary. We propose an adaptive sieve estimator for the span of the time varying loading matrix of a locally stationary factor process. A uniformly consistent estimator of the effective number of factors is developed via eigenanalysis of a non-negative definite time varying matrix. We also propose a possibly high-dimensional bootstrap test for the hypothesis of constant factor loadings by comparing the kernels of the covariance matrices of the whole time series with their local counterparts. This test avoids the assumption that factors and idiosyncratic errors are stationary or the covariance matrix of factors is time-invariant. Our results cover both the case of white noise idiosyncratic errors and the case of serially correlated idiosyncratic errors. We examine the finite sample performance of our proposed estimator and test via simulation studies and real data analysis.
Abstract：
Since the seminal work about social effect identification by Manski (1993, 1995, 2000), the econometric literature of social interactions has been paying considerable attention to separately identifying and consistently estimating three types of effects: endogenous social effects, exogenous (or contextual) social effects, and correlated effects. Besides the identification problem, we argue that model selection problem is particularly important in applying social interaction models. For a given social interaction model, there could be many equivalent or arbitrarily close social interaction models, in terms of prediction error. Also, it is possible to obtain spurious social effects in a system of independent equations. In this paper, we establishes consistent nonlinear least squares (NLS) estimators and predictive-efficient model selection (MS) criteria for a finite set of parametric, potentially misspecified, and non-nested social interaction models (with either observed or unobserved networks). The models under consideration include various forms of linear and nonlinear regression and spatial models. The new NLS estimators are required for deriving MS criteria from prediction errors and for proving asymptotic results. The MS criteria are asymptotically efficient whether the true model is included or not. A discussion on ``true model'', parsimony principle, and MS consistency is provided.
Abstract:
We propose a unified model averaging (MA) approach for a broad class of forecasting targets. This approach is established by minimizing an asymptotic risk based on the expected Bregman divergence of a combined forecast, relative to the optimal forecast of the forecasting target, under local(-to-zero) asymptotics. It can be flexibly applied to develop effective MA methods across various forecasting contexts, including but not limited to univariate and multivariate mean forecasting, volatility forecasting, probabilistic forecasting, and density forecasting. As illustrative examples, we present a series of simulation experiments and empirical cases that demonstrate strong numerical performance of our approach in forecasting. (This is joint work with Yi-Ting Chen and Chu-An Liu.)
Abstract:
In causal inference, one of the most widely used—but fundamentally untestable—assumptions is unconfoundedness. It allows us to estimate treatment effects by assuming that, once we control for observable variables, treatment assignment is essentially random. But because we can never observe all potential outcomes, this assumption is impossible to verify. To address this limitation, I explore a nonparametric relaxation of unconfoundedness called the conditional moment constraint. Instead of relying on full independence, this approach looks at differences between conditional expectations with and without treatment conditioning. These constraints allow us to derive bounds—rather than precise estimates—for both the Conditional Average Treatment Effect (CATE) and the Average Treatment Effect (ATE).