Research Interests

Working Papers

  • Current: Theoretical and Applied Econometrics; Empirical Asset Pricing; Financial Econometrics

  • More: Big Data Analysis; Financial Machine Learning; FinTech

       This paper investigates the impact of firm networks on asset pricing using rich data and conditional factor pricing models. Drawing on supplier-customer network data for U.S. public firms, a comprehensive dataset on stock returns and firm characteristics, and the Instrumented Principal Component Analysis (IPCA) approach, this study provides novel evidence that the historical stock returns of a firm's major customers consistently capture a new aggregate risk, even when controlling for numerous firm-specific characteristics. Moreover, both a firm’s network centrality and its customers’ historical returns significantly influence its exposure to existing risk factors. These findings have important implications for investment decisions. A long-short strategy based on model-implied expected returns from a network-augmented conditional factor pricing model—whether using latent or observable factors—delivers significantly higher out-of-sample excess returns and Sharpe ratios than models without network augmentation. This holds true even though the latter models have already incorporated a broad range of firm characteristics.

       This paper considers a general multi-level group factor model in which three different types of factors influence different groups of cross-sectional entities: local factors, relevant to a single group; regional factors, common to proper subsets of groups; and global factors, influencing all groups. For the model, we develop a new projection-based approach for estimating the global factors, and propose two novel methods: the "Projection-based Modified Eigenvalue Ratio" (ProjectionMER) and the "Projection-based Eigenvalue Thresholding" (ProjectionET), which consistently estimate the number of global factors. The proposed methods outperform existing approaches in the literature in finite samples. By sequentially applying these methods across various subsets, we can estimate the local, regional and global factors without specifying the regional affiliations of individual groups. Our Monte Carlo simulation results and topical empirical results provide promising evidence for the applicability of these methods.

       Panel data models with cross-sectionally heteroskedastic data often suffer from the well-known incidental parameters problem. Some recent studies have proposed that the structural parameters (common parameters to all of the cross-sectional entities) can be consistently estimated if they are estimated jointly with the cross-sectionally weighted averages of the incidental parameters. In this paper, we provide a sufficient condition under which the proposed methods can yield consistent estimators of the structural parameters and the consistently estimated asymptotic variance-covariance matrix of the structural parameter estimators. With the condition, we show that the unrestricted factor IV method proposed by Robertson and Sarafidis (2015, Journal of Econometrics) and the transformed likelihood method of Hayakawa and Pesaran (2015, Journal of Econometrics) can produce consistent estimates for the panel data models with unknown common factors or dynamic panel models with fixed individual entity-specific effects.

Pre-Doctoral Publication

Short-Term Exchange Rate Predictability, with Yu Ren and Qin Wang, Finance Research Letters 28 (2019): 148-152.