About

I am a PhD Candidate in the Research School of Economics at Australian National University. My research interests are Markov decision process, economic networks, and economic computation. Here is my Curriculum Vitae.

Working Papers

Dynamic Programming with State-Action-Dependent Discounting

In this paper, we extend the discrete-time dynamic programming to the case of state-action-dependent discounting. We establish a sufficient condition known as “eventual discounting” to guarantee the standard optimality results. The condition becomes necessary for the existence of policy value when state space is compact. Our research encompasses dynamic programming with both bounded and unbounded rewards. Furthermore, we extend the scope of eventual discounting to applications involving risk-sensitive preferences.

Temporal-Difference Learning with State-Action-Dependent Discounting

This paper extends model-free learning algorithms, including Q-learning, SARSA, and double Q-learning to the learning with state-action-dependent discount factors. We allow the discount factor to be greater than one with positive probability, but the expected multiplicative of discount factors satisfies the “eventually discounting” condition in the sense that we replace $β < 1$ by $ρ(L) < 1$, where $ρ(L)$ denotes the spectral radius of an appropriate matrix dominating the expected discounted future value.

Uniqueness of Equilibria in Interactive Networks

We study a unified network framework and show that equilibrium exists and is unique (almost surely) under either eventually contracting or non-expansive assumptions. We also discuss the computation methods and show the necessity of boundedness conditions when interaction functions are non-expansive. Applying the equilibrium in the study of systemic risk, we provide a measure to determine the key player who causes the most significant impact if removed from the network.

Explaining Systematic Departures from Gibrat’s Law

One of the most well-known ”laws” in economics is Gibrat’s Law of proportionate effect. Concerning firm dynamics, Gibrat’s law predicts that firm growth will be independent of its size. While Gibrat’s Law is frequently used as a benchmark in models of industry dynamics, the majority of recent empirical literature demonstrates a systematic departure from Gibrat’s Law: small firms exhibit higher volatility than larger ones. This paper aims to provide an explanation for this systematic deviation. Utilizing a hierarchical production network, we demonstrate that upstream firms, which tend to be smaller, display more volatility than downstream firms. The model also offers an explanation for the observation that the firm size distribution follows a power law.

Other Projects

Forecasting project
This is a project for EMET8012 course, which I completed during my master’s studies.
Zurcher as a Q-learner
Power and Size Analysis in Two-Stage Least Squares (TSLS) Models with Weak Instruments This is a project for EMET8002 course. Abstract: In instrumental variables (IV) regression, particularly using Two-Stage Least Squares (TSLS), the reliability of traditional inference methods can be compromised when instruments are weak—meaning they exhibit a low correlation with the endogenous explanatory variables. This project investigates the distribution of IV estimators under such conditions and examines the performance of various tests, particularly the Anderson-Rubin (AR) test.
Analyzing Herfindahl Index for Pareto and Log-Normal Distributions This project compares Pareto and log-normal distributions under Zipf’s law by analyzing the Herfindahl index, a measure of market concentration. The goal was to understand how well the log-normal distribution, when adjusted, matches the Pareto distribution in terms of the Herfindahl index and other statistical properties.