I am a Ph.D. candidate in Economics at Indiana University Bloomington (expected May 2027),
specializing in econometrics and machine learning.
My research develops statistical and machine-learning methods for high-dimensional and nonparametric econometric models.
Job Market Paper
Random Forests for Nonsmooth Binary Choice Models
I develop a semiparametric estimation procedure for binary choice models using random forests without imposing smoothness or a parametric functional form on the systematic component.
We propose a semiparametric estimation procedure for binary choice models using random forests without imposing smoothness or a parametric functional form on the systematic component. Our method uses existing random forest implementations to obtain the tree partitions without additional numerical optimization at each split. The proposed estimator then transforms the adjusted within-leaf sample proportions through the inverse of the known error distribution function and averages the resulting tree-level estimates of the systematic component. We use cross-validation to jointly selects the standard random forest tuning parameters and the additional tuning parameter introduced by our procedure.
Under suitable regularity conditions, the theoretical analysis establishes L2 convergence rates while accommodating continuous and discrete covariates whose dimensions may grow polynomially with the sample size. The simulation results support adapting random forests to the binary choice structure. Compared with the benchmark methods, our proposed estimator performs particularly well when the systematic component exhibits sharp nonlinearities or discontinuities. An empirical application to P2P lending market data demonstrates how the estimator can capture heterogeneous and nonlinear effects on the willingness to default without imposing prespecified credit score categories.
Local projection (LP) has become a popular tool for estimating impulse responses in macroeconomics. In this paper, we focus on stationary processes whose impulse responses decay too slowly to be absolutely summable. The LP method is particularly attractive in this context because it does not require explicitly modeling or estimating the parameters governing the slow decay. However, existing asymptotic theory for the semiparametric LP regression with a growing lag order does not cover this setting, because of the failure of the commonly assumed condition on the tail sum of the coefficients. A different theory is therefore needed. In a multivariate system, we establish consistency and a convergence-rate bound for the LP estimator, allowing for both growing propagation horizons and high-order dependence in the shocks —properties that have not been formally established for other semiparametric impulse-response estimators. We further extend our results to the structural LP estimator, allowing the Wold coefficients to be not absolutely summable.
Estimation of Large Dimensional Binary Choice Models with Penalized Random Forests
We propose a modified random forest estimator for binary choice models that uses the negative log-likelihood function in place of the impurity measures employed in standard random forests. In addition, we incorporate both L1 and L2 penalties into the recursive partitioning procedure to regularize the coefficients associated with the systematic component. Simulation studies suggest that the proposed estimator performs well in finite samples and is particularly effective at estimating discontinuous or nearly discontinuous functions.
Asymptotic Theory for Nonparametric Regression by Random Forest
with Joon Y. Park and Doyeon Pyun
In this paper, we establish consistency results for nonparametric regression by random forests. Based on the work of Chi et al. (2022), we provide several theoretical advancements, including (i) dispensing with the symmetrically distributed error assumption, (ii) allowing for dependence between the error and the covariates, (iii) accommodating discrete-type covariates, and (iv) improving convergence rates. Moreover, using a regularization method for balanced splits, we demonstrate that convergence rates can be further improved by reducing the variance part of the mean-squared error of the estimator.
Inference of High Dimensional Local Projection under General Conditions
with Keli Xu
The project aims to extend inference for local projections to high-dimensional settings while maintaining a set of weak assumptions. The asymptotic normality for estimated impulse responses is derived with an estimable variance.
Zhang, J. (2021). Journal of Risk and Financial Management, 14(3), 93.
Research completed before my Ph.D. studies.
This study finds an N-shaped long-run relationship between CO₂ emissions and real GDP per capita in China, alongside a positive effect of energy consumption and a negative effect of urbanization.
Music has been a significant part of my life. I started teaching myself piano and composition when I was 14, and since then, I’ve enjoyed creating my own pieces and performing classical music in my free time.