Hi all, we have a new research paper titled “Shadow Banking in a Crisis: Evidence from FinTech During COVID-19” forthcoming in the Journal of Financial and Quantitative Analysis (JFQA).* Compared with traditional banks, shadow banks are more likely to expand credit access to new and financially constrained borrowers, but they face a bigger challenge in maintaining sustainable delinquency rates. Use the link Shadow Banking in a Crisis: Evidence from FinTech During COVID-19 by Zhengyang Bao, Difang Huang :: SSRN
Hello @Difang_Huang_Monash_University Super interesting paper — and outside my area of expertise, but I hope my questions aren’t too off base. No urgency to answer, but had a few questions:
- You focused on differences in differences specifications, and use matching estimates as a robustness check. Was there a particular reason you chose this analysis strategy as opposed to using a mixed design with both diff in diff and matching or focusing primarily on the matching design?
- The study had an impressively large sample size, revealing some significant but very small magnitude effect sizes (e.g. a 0.2% increase in the loan rate in Wuhan) (other things had large effect sizes, such as the five times higher delinquency rates between banks and fintech firms). Given a sufficiently large sample size almost any relationship or variable could become statistically significant. How do you think about the importance or implications of the small-magnitude effect sizes?
- With regard to the matching, the paper used a propensity score and entropy balancing approach. I’m not first-hand experienced with this, but Gary King has argued that propensity score matching among matching methods is relatively weak (see here, he has gone so far as to argue that propensity scores should not be used in matching). In the present case, given the large n characteristics of the sample, coarsened exact matching or even simple exact matching would not have their normal draw backs of needing large sample sizes to maintain statistical power. Did you consider other matching methods, such as coarsened exact matching or genetic matching which may provide improved match balance?
Thanks again for the fascinating paper!
Thank you very much @Ryan_Fox_Squire_SafeGraph
- We use multiple econometric specifications including the difference-in-differences and matching method for robustness. All specifications yield consistent results for empirical analysis.
- I completely agree with you that given large sample size, almost any relationship or variable could become statistically significant. The effect in more COVID affected area including Wuhan is statistically significant while small in magnitude, showing compared with the less affected area, the delinquency rate for borrowers in more affected areas are slightly higher and the severity of pandemic may not fully explain our main empirical results.
- Thank you very much for providing the most recent discussion on propensity score matching method. We may consider other matching method to probe the further robustness of our results.
@Difang_Huang_Monash_University Interesting study. I’m not too surprised, although I’d be curious if when you update the study, the subsequent boom shows up stronger for the FinTech group. On Table 3: The Fintech indicator for loan amount is 2132, the cycle duration is -4.3, and the interest rate is 2.6. The cycle duration and interest rate makes sense, although I’m surprised the interest rate is so much higher. Did you do some winsorizing to see how sensitive these results are? The loan amount seems odd to me, especially that it gets bigger for part (b) of Table 3. Is this always consistent? I’m real curious if Fintech shows up as more profitable after the pandemic because of their willingness to take greater risks.
Thank you very much for your comments @Thomas_Young_Econometric_Studios_Utah_Legislature. Sorry for my late reply as I just saw your comments. As there are new policy in FinTech industry enacted in July 2020, we have to stopped our data sample until June 2020 and may not show the subsequent boom for FinTech groups. As for the results in Table 3, we don’t winsorize sample while the results are still robust if we do winsorizing.