Abstract: Traditional hard clustering methods fail to capture the complex, overlapping risk linkages present in modern markets. While recent work (Ding et al., 2024) identifies broad block structures in stock co-jumps, we demonstrate that this framework can be significantly enhanced by incorporating mixed membership. Specifically, we propose a Degree Corrected Mixed Membership network model with Dependent Multivariate Poisson edges that allows stocks to belong to multiple latent risk communities simultaneously and a corresponding estimation algorithm. We prove that our estimator is asymptotically consistent. Empirically, we show that: (1) the mixed membership framework reveals a more informative and granular risk structure than rigid classifications; (2) peer groups defined by our model generate momentum strategies that significantly outperform benchmarks such as selfmomentum, or momentum based on GICS or analyst coverage; and (3) a \"purity\" measure derived from our model predicts stock volatility and Sharpe ratios, offering a new dimension for risk profiling.
Bio:\u00a0Yingying Li is Fung Term Professor of Business (\u99ae\u6c0f\u5546\u5b78\u6559\u6388) and Chair Professor at the Department of Information System, Business Statistics and Operations Management (ISOM) and the Department of Finance at Hong Kong University of Science and Technology (HKUST). She served as acting head of the FinTech Thrust at HKUST(GZ) at its initial stage. Before joining HKUST, she also held positions as lecturer and postdoctoral fellow at the Bendheim Center for Finance and the Operations Research and Financial Engineering department at Princeton University. Dr. Li received her Ph. D in Statistics from the University of Chicago.
Prof. Li's research focuses on statistical learning for financial big data, large portfolio optimization, individualized financial decision making, volatility estimation and prediction etc. Prof. Li has published in top journals in statistics, finance and economics, such as Econometrica, Review of Financial Studies, Journal of Financial Economics, Annals of Statistics, Journal of American Statistical Association, and Journal of Econometrics.
Prof. Li is an RGC Senior Research Fellow, elected Fellow of the Society for Financial Econometrics (SoFiE), Fellow of Journal of Econometrics, NSFC YSF Cat-A awardee, and NSFC Excellent Young Scientist. She served or is serving as an Associate Editor for the Journal of American Statistical Association, Management Science, Journal of Econometrics, Journal of Business & Economic Statistics and Journal of Financial Econometrics; a council member for the Society for Financial Econometrics, and a panel member of the Hong Kong Research Grants Council (RGC) and the University Grants Committee (UGC).
Abstract: Large-scale causal mediation, replication, and pleiotropy analyses may appear to be unrelated problems. However, they all involve testing a large number of composite null hypotheses, a statistically challenging task arising in many scientific settings, especially in genomic and biobank studies. A composite null is defined as the union of K individual null hypotheses, and is rejected only when all individual hypotheses are false. We will first show that classical tests such as Wald and likelihood ratios fail to follow chi-square tests and are often overly conservative, leading to substantial power loss. \u00a0We propose several approaches to address this problem. We show that the divide-aggregate composite-null test (DACT) \u00a0by leveraging a large number of tests and Efron\u2019s empirical null framework works well when signals are weak and sparse, but can exhibit error rate inflation in the presence of dense and strong signals. To overcome this limitation, we propose \u00a0a tail likelihood ratio (TLR)\u00a0 test that is robust and powerful across diverse settings, encompassing weak to strong signal strengths, sparse to dense signal proportions, symmetric and unsymmetric signal distributions, and correlations among the large number of tests. \u00a0We show that local FDR for a large number of composite nulls can be expressed as a function of a likelihood ratio, which is approximated in the tail region to construct the TLR statistic. We derive the null distribution of the TLR statistic by leveraging information from the large number of tests, and account for correlation among tests using a principal component-based approximation. Extensive simulations demonstrate that the proposed method maintains accurate error rate control in multiple comparisons and exhibits superior power gains compared to existing methods. \u00a0Applications to multiple examples further illustrate the performance and utility of the proposed framework.\u00a0
\u00a0
Abstract: Integration of statistics and generative AI plays a pivotal role for accelerating trustworthy cross-domain scientific discovery. Recent advances in generative models have dramatically increased the availability and use of synthetic data across scientific domains. While these developments create exciting opportunities for empowering data analysis, they also raise fundamental statistical challenges regarding how synthetic data can be used in a valid, reliable, and principled manner.\u00a0In this talk, we first discuss the current landscape of high-fidelity high-dimensional synthetic data generation using generative AI models such as transformer, \u00a0diffusion, and flow matching based models. More importantly, we present a principled framework for incorporating synthetic data in downstream statistical analysis that ensures valid statistical inference \u00a0even when generative AI models are misspecified. We show that the proposed synthetic data assisted and augmented methods for continuous and discreate outcomes that integrate observed and synthetic data are robust to misspecified black-box generative models and can improve statistical inferential power when the generative AI models are informative. \u00a0We demonstrate the utility of these synthetic data assisted methods to the analysis of the UK biobank data, by performing genome-wide association studies (GWAS) of proteomic data and whole-genome sequencing (WGS) analyses of brain imaging phenotypes, and binary mental health survey data both characterized by substantial missingness (about 70-90%).
\u00a0