Research 

Forecasting Loan Risk of Banks with Machine Learning in Main Street Lending Program (Job Market paper)

The Main Street Lending Program (MSLP) aims to facilitate banks’ lending to small and medium-sized businesses to recover from the COVID-19 pandemic, designed by the Federal Reserve and assets backed by the Treasury Department. However, banks are reluctant to lend due to uncertainties about firms’ repayment capabilities. Accordingly, we use machine- learning algorithms to forecast the bank loan risk premium, with vintages of the Main Street Lending Program, bank, and macroeconomic predictors. We compare the predictive ability of the linear and machine learning models. The results indicate that Support Vector Machines(SVM) have the highest accuracy. We also analyze the feature importance using the Shapley Additive Explanations(SHAP) value. We find that the Transaction Fee Rate deter- mined by the Federal Reserve is the second important variable for forecasting, and policy- makers can improve the policy through this channel, while undoubtedly, the macroeconomic environment, such as CPI contributes the most. The evidence suggests that the MSLP could serve as a form of “forward guidance” that the Federal Reserve could influence banks’ expectations and accelerate the process of allocating funds from the Fed to businesses. 

Explainable EU ETS Prices Forecasting Driven by Multi-source Heterogeneous Data: A Mixed-Frequency Perspective (2024). R&R by Applied Energy (joint with Zhenpeng Tang, Dinggao Liu, Kaijie Chen, Yi Cai)

The EU Emissions Trading System (EU ETS) is the world’s largest carbon market and plays a crucial role in helping the EU achieve its carbon neutrality goals. However, in the era of big data, predicting the EU Allowance (EUA) prices using multimodal data and addressing mixed-frequency issues remain challenging. In this paper, we propose a novel interpretable multi- source heterogeneous data-driven framework for predicting the prices. This forecasting system integrates unstructured data (online news and search trends) with structured market-related variables, and employs text mining techniques to enhance the prediction accuracy. To address the issue of mixed-frequency data from heterogeneous sources, we apply the deep generative simulation method Time-series Generative Adversarial Networks (TimeGAN) to generate high-frequency data from low-frequency inputs. We utilize the Temporal Fusion Transformer (TFT) model to predict the EUA prices, which provides interpretable results that are both global and time-varying. Empirical results demonstrate that the proposed framework outperforms benchmark models across multiple predictive evaluation criteria. It achieves a minimum improvement of 12% and a maximum of 30.3% in directional accuracy. This study provides robust forecasting tools for policymakers and investors in the EU ETS, offering critical information to more effectively address market fluctuations and furnishing a more scientific basis for risk management and investment decisions.

How Birth Order and Gender Affect the Test Score of Children in India

The social structure of India often emphasizes the importance of first-born sons and the relatively lower importance of daughters. This research explores how gender and birth order affect children’s test scores from 5 to 16 years old in rural India, particularly whether the first boy in the family receives differential educational investment relative to other children in the household. This paper analyzes the influence of gender, age, and birth order on children’s test scores in rural India by utilizing an ordinary least squares (OLS) regression model, controlling for age and birth order, on a dataset of over half a million test scores from multiple subjects of children aged 5 to 16. The result shows that 1) birth order and gender both affect children’s test scores; 2) schooling helps eliminate the test score difference caused by birth order for boys but not for girls; 3) first-born girls dominate younger girls when they are at school; 4) first-born boys have the negative effect on later boys before 15, but they generally have the positive effect on later girls. This paper analyzes the birth order and gender from a different perspective that adds the factor of school dropouts. It also considers the difference between later boys’ and later girls’ education conditional on whether they have a big brother. 

How the Main Street Lending Program affects banks’ profitability during the COVID-19 Pandemic 

This paper explores the effect of the Main Street Lending Program on bank’s profitability. The Main Street Lending Program lends money to small and medium-sized firms through banks. However, we do not know whether banks will be harmed or helped during the lending program from the perspective of profitability. I use bank efficiency (total cost-to-income ratio), net income, net interest income, and non-interest income to measure profitability. I use PSM-DID, synthetic-DID, and IV to measure the causality and find that the program truly has negative effect on bank profitability in the way of net income, net interest income, and non-interest income. I use these methods to guarantee the parallel trend assumption and IV strategy is to address the possible endogeneity problem that cannot be solved by DID. 

Does the Dodd-Frank Act Affect Banks’ Systematic Risks during the COVID-19 Pandemic

This study investigates the impact of the Dodd-Frank Act on the systematic risk of banks  during the COVID-19 pandemic. Systematic risk is quantified using monthly beta values, and the Difference-in-Differences-in-Differences (DDD) method is employed to examine the causal effect. The findings indicate that banks compliant with the Act experienced a significant reduction in systematic risk during the pandemic. The Dodd-Frank Act functions as a "simulation shock," enhancing banks' resilience to risk and their capacity to withstand adverse conditions.