Yizhou Kuang (匡逸舟)
I joined the Department of Economics at the University of Manchester as a Lecturer (Assistant Professor) in 2023, following my graduation from Cornell University. My research focuses on macroeconometrics. I am also organizing the brown bag seminars at Manchester.
Outside of work, my enthusiasm lies in rock climbing, especially bouldering. If you share this passion, I welcome the opportunity to climb together.
WHAT I DO
Research
Working Papers
Bayesian Sensitivity Analysis for Set-identified Structural Models Partial Identification Bayesian Policy Analysis
R&R at Journal of Business & Economic Statistics
PDF
Abstract: This paper proposes a new algorithm to conduct robust Bayesian analysis for set-identified structural models. It combines standard Bayesian procedure with a characterization of observationally equivalent parameters. The algorithm finds the range of posterior means and the Bayesian credible region of both the structural parameters and any parameters of interest, ensuring robustness against the selection of priors within a class that produces identical marginal likelihoods. I provide theoretical support for this algorithm and apply the method to several monetary policy models, to show its relevance in policy analysis. The methodology finds that, in set-identified models, parameters of primary interest like impulse responses can have very different implications based on the prior, even within the same class. Additionally, optimal monetary policy rules could vary with the choice of prior within that class, particularly when historical policy parameters are not identified.
R&R at Journal of Business & Economic Statistics
Abstract: This paper proposes a new algorithm to conduct robust Bayesian analysis for set-identified structural models. It combines standard Bayesian procedure with a characterization of observationally equivalent parameters. The algorithm finds the range of posterior means and the Bayesian credible region of both the structural parameters and any parameters of interest, ensuring robustness against the selection of priors within a class that produces identical marginal likelihoods. I provide theoretical support for this algorithm and apply the method to several monetary policy models, to show its relevance in policy analysis. The methodology finds that, in set-identified models, parameters of primary interest like impulse responses can have very different implications based on the prior, even within the same class. Additionally, optimal monetary policy rules could vary with the choice of prior within that class, particularly when historical policy parameters are not identified.
Individual and Common Information: Model-free Evidence from Probability Forecasts Information Acquisition Bayesian Policy Analysis
with Kristoffer Nimark, R&R at Journal of Applied Econometrics
PDF Slides Codes
Abstract: We propose a method to empirically decompose a cross-section of observed belief revisions into components driven by individual and common information under weak assumptions. We define a common signal as the single signal that if observed by all agents can explain the maximum amount of belief revisions across agents. Individual signals are defined to explain the residual belief revisions unaccounted for by the common signal. When applied to probability forecasts from the Survey of Professional Forecasters we find that individual signals account for more of the observed belief revisions than common signals. There is a large cross-sectional heterogeneity in signal precision, with about 2/3 of forecasters observing individual signals that are more precise than the common signal. Unconditionally, the informativeness of individual and common signals are positively correlated. Inflation volatility, perceived stock market volatility and a high risk of recession are all factors associated with increased informativeness and precision of both individual and common signals. We discuss the implications of our findings for theoretical models of information acquisition and we show how our procedure maps into alternative information structures.
with Kristoffer Nimark, R&R at Journal of Applied Econometrics
PDF Slides Codes
Abstract: We propose a method to empirically decompose a cross-section of observed belief revisions into components driven by individual and common information under weak assumptions. We define a common signal as the single signal that if observed by all agents can explain the maximum amount of belief revisions across agents. Individual signals are defined to explain the residual belief revisions unaccounted for by the common signal. When applied to probability forecasts from the Survey of Professional Forecasters we find that individual signals account for more of the observed belief revisions than common signals. There is a large cross-sectional heterogeneity in signal precision, with about 2/3 of forecasters observing individual signals that are more precise than the common signal. Unconditionally, the informativeness of individual and common signals are positively correlated. Inflation volatility, perceived stock market volatility and a high risk of recession are all factors associated with increased informativeness and precision of both individual and common signals. We discuss the implications of our findings for theoretical models of information acquisition and we show how our procedure maps into alternative information structures.
Identification-aware Markov Chain Monte Carlo Partial Identification Bayesian
with Toru Kitagawa
PDF Slides
Abstract: Leaving posterior sensitivity concerns aside, non-identifiability of the parameters does not raise a difficulty for Bayesian inference as far as the posterior is proper, but multi-modality or flat regions of the posterior induced by the lack of identification leaves a challenge for modern Bayesian computation. Sampling methods often struggle with slow or non-convergence when dealing with multiple modes or flat regions of the target distributions. This paper develops a novel Markov chain Monte Carlo (MCMC) approach for non-identified models, leveraging the knowledge of observationally equivalent sets of parameters, and highlights an important role that identification plays in Bayesian analysis. We show that our proposal overcomes the issues of being trapped in a local mode and achieves a faster rate of convergence than the existing MCMC techniques including random walk Metropolis-Hastings and Hamiltonian Monte Carlo. The gain in the speed of convergence is more significant as the dimension or cardinality of the identified sets increases. Simulation studies show its superior performance compared to other popular computational methods including sequential Monte Carlo. We illustrate use and effectiveness of our proposal in a non-invertible structural vector moving average (SVMA) model.
with Toru Kitagawa
PDF Slides
Abstract: Leaving posterior sensitivity concerns aside, non-identifiability of the parameters does not raise a difficulty for Bayesian inference as far as the posterior is proper, but multi-modality or flat regions of the posterior induced by the lack of identification leaves a challenge for modern Bayesian computation. Sampling methods often struggle with slow or non-convergence when dealing with multiple modes or flat regions of the target distributions. This paper develops a novel Markov chain Monte Carlo (MCMC) approach for non-identified models, leveraging the knowledge of observationally equivalent sets of parameters, and highlights an important role that identification plays in Bayesian analysis. We show that our proposal overcomes the issues of being trapped in a local mode and achieves a faster rate of convergence than the existing MCMC techniques including random walk Metropolis-Hastings and Hamiltonian Monte Carlo. The gain in the speed of convergence is more significant as the dimension or cardinality of the identified sets increases. Simulation studies show its superior performance compared to other popular computational methods including sequential Monte Carlo. We illustrate use and effectiveness of our proposal in a non-invertible structural vector moving average (SVMA) model.
Outlier-robust BVARs Bayesian
with Dimitris Korobilis
PDF Slides
Abstract: We develop Bayesian vector autoregressions (VARs) that are robust to outliers, drawing on ideas from empirical risk minimization rather than on detecting, deleting, or parametrically modeling outliers. Robustness is achieved from two complementary perspectives, a loss-based approach that replaces the Gaussian likelihood with a robust loss inside a generalized Bayesian update, and a conventional likelihood-based approach that embeds a smooth robust loss in an exact scale-mixture likelihood. Both approaches to Bayesian VAR inference discount aberrant observations automatically, and each comes with a dedicated, efficient estimation algorithm. We prove that robust loss learning bounds the influence of an outlier on coefficients and forecasts, whereas its effect under Gaussian learning grows with the size of the outlier. Forecasting U.S. macroeconomic variables through 2025Q2, the robust VARs cut one-quarter-ahead forecast errors by 16 to 18 percent during the COVID19 period at negligible cost in tranquil periods.
with Dimitris Korobilis
PDF Slides
Abstract: We develop Bayesian vector autoregressions (VARs) that are robust to outliers, drawing on ideas from empirical risk minimization rather than on detecting, deleting, or parametrically modeling outliers. Robustness is achieved from two complementary perspectives, a loss-based approach that replaces the Gaussian likelihood with a robust loss inside a generalized Bayesian update, and a conventional likelihood-based approach that embeds a smooth robust loss in an exact scale-mixture likelihood. Both approaches to Bayesian VAR inference discount aberrant observations automatically, and each comes with a dedicated, efficient estimation algorithm. We prove that robust loss learning bounds the influence of an outlier on coefficients and forecasts, whereas its effect under Gaussian learning grows with the size of the outlier. Forecasting U.S. macroeconomic variables through 2025Q2, the robust VARs cut one-quarter-ahead forecast errors by 16 to 18 percent during the COVID19 period at negligible cost in tranquil periods.
Revision Risk in Real-Time Macroeconomic Forecasting Forecasting Partial Identification Policy Analysis
Submitted
PDF Slides
Abstract: Macroeconomic outcomes are revised after release, so the same forecast can be evaluated against several official values. These revisions affect forecast evaluation and uncertainty. Holding a forecast fixed, I show how its loss and ranking relative to another forecast can change across releases. I also develop prediction intervals for a specified official release using information available when the forecast is issued. Separate histories identify the distributions of early errors and revisions, but not their dependence or the distribution of their sum. I characterize the resulting pointwise Fréchet-Makarov bounds and derive the shortest worst-case prediction interval for fixed empirical distributions. For SPF median forecasts, revisions through roughly 180 days average 8.2 percent of later-release MSE for real activity and 3.6 percent for inflation. In the out-of-sample evaluation, SPF intervals undercover during COVID. Separately, national-account results favor direct or revision-aware calibration when later-release error histories are informative.
Submitted
PDF Slides
Abstract: Macroeconomic outcomes are revised after release, so the same forecast can be evaluated against several official values. These revisions affect forecast evaluation and uncertainty. Holding a forecast fixed, I show how its loss and ranking relative to another forecast can change across releases. I also develop prediction intervals for a specified official release using information available when the forecast is issued. Separate histories identify the distributions of early errors and revisions, but not their dependence or the distribution of their sum. I characterize the resulting pointwise Fréchet-Makarov bounds and derive the shortest worst-case prediction interval for fixed empirical distributions. For SPF median forecasts, revisions through roughly 180 days average 8.2 percent of later-release MSE for real activity and 3.6 percent for inflation. In the out-of-sample evaluation, SPF intervals undercover during COVID. Separately, national-account results favor direct or revision-aware calibration when later-release error histories are informative.
Centralized or Decentralized? An Empirical Model on Task Assignment of Government in Pandemics IO Policy Analysis
with Qiwei He
PDF Slides
Abstract: Mitigation policy in a federation trades a delayed, cross-border health benefit against immediate local employment and political costs. We estimate a dynamic game among seven Northeastern U.S. states during the first wave of COVID-19, combining a death-targeted epidemic transition with cross-state mobility exposure and forward-simulation inequalities that recover government payoffs. Mortality is costly everywhere, running at 45 percent of the employment burden at first-wave death rates. The cost of changing policy dominates both, and it is directional: loosening by ten points of the stringency index while deaths are still rising costs the equivalent of a 56-point employment gap for that day, against 18 points for the same tightening after the peak. That asymmetry, rather than the level of political pressure a state faces, is what holds the observed policy path flat. Centralizing policy within the estimated rule class lowers social cost by 2.3 percent when one common reaction function is imposed on all seven states, and by 5.4 percent when the reaction is allowed to differ across them; a national planner and a president select the same rule. Most of what centralization buys is therefore the freedom to treat states differently, not coordination itself.
with Qiwei He
PDF Slides
Abstract: Mitigation policy in a federation trades a delayed, cross-border health benefit against immediate local employment and political costs. We estimate a dynamic game among seven Northeastern U.S. states during the first wave of COVID-19, combining a death-targeted epidemic transition with cross-state mobility exposure and forward-simulation inequalities that recover government payoffs. Mortality is costly everywhere, running at 45 percent of the employment burden at first-wave death rates. The cost of changing policy dominates both, and it is directional: loosening by ten points of the stringency index while deaths are still rising costs the equivalent of a 56-point employment gap for that day, against 18 points for the same tightening after the peak. That asymmetry, rather than the level of political pressure a state faces, is what holds the observed policy path flat. Centralizing policy within the estimated rule class lowers social cost by 2.3 percent when one common reaction function is imposed on all seven states, and by 5.4 percent when the reaction is allowed to differ across them; a national planner and a president select the same rule. Most of what centralization buys is therefore the freedom to treat states differently, not coordination itself.
Work in Progress
Supervised MIDAS Latent-Factor Regression Time Series ML
with Rosnel Sessinou
PDF Slides
Abstract: not yet available
with Rosnel Sessinou
PDF Slides
Abstract: not yet available
Dormant Projects
Venture Capital Investment Geography Search & Matching IO
with Qinshu Xue, Bin Zhao
PDF Slides
Abstract: This paper studies the geographic concentration of VC investment in the US. We discovered that both VC firms and VC-backed companies are highly clustered in the Bay-Boston-NY area. Using the detailed Uniform Commercial Code (UCC) filings that are self-reported by lenders to stake a claim to specific pieces of collateral, we track the transaction of capital across firms. We propose the vintage capital market density as an essential determinant of VC investment concentration. Since young firms can benefit from cheap vintage capital while old firms can exit with a high scrap value, VC investments are attracted due to a high entry rate and a low exit cost. By modeling a market for the vintage capital, we aim to endogenize the scrap value of firms. Our paper highlights the critical role of the capital market in determining the industry and venture capital agglomeration. Industry policy that helps promote local capital market density would also attract VC investment in places with the greatest economic need.
with Qinshu Xue, Bin Zhao
PDF Slides
Abstract: This paper studies the geographic concentration of VC investment in the US. We discovered that both VC firms and VC-backed companies are highly clustered in the Bay-Boston-NY area. Using the detailed Uniform Commercial Code (UCC) filings that are self-reported by lenders to stake a claim to specific pieces of collateral, we track the transaction of capital across firms. We propose the vintage capital market density as an essential determinant of VC investment concentration. Since young firms can benefit from cheap vintage capital while old firms can exit with a high scrap value, VC investments are attracted due to a high entry rate and a low exit cost. By modeling a market for the vintage capital, we aim to endogenize the scrap value of firms. Our paper highlights the critical role of the capital market in determining the industry and venture capital agglomeration. Industry policy that helps promote local capital market density would also attract VC investment in places with the greatest economic need.
Nowcasting with Dynamic Factors: A Penalized Model Averaging Approach Time Series ML
with Yongmiao Hong, Yuying Sun
PDF Slides
Abstract: With the advent of complex information systems to collect data, real-time nowcasting faces various challenges, including a more significant number of predictors, higher order of lags, unbalanced data structure, model uncertainty and complexity in bridging high-frequency information contained. To address these issues, this paper proposes a new real-time nowcasting forecast combination with dynamic factor regressions, which deletes redundant predictors and simultaneously selects optimal weights for candidate models. We show that the selected weight achieves asymptotic optimality and consistency, even when all candidate models are misspecified. The proposed estimator is consistent and asymptotically Gaussian if the true model is included in candidate models. Simulation results show that the proposed method yields lower mean square forecast errors than alternative nowcasting methods, including MIDAS in Ghysels et al. (2004), GARS in Giannone et al. (2008), and FADL-MIDAS in Andreou et al. (2013). The proposed method is applied to forecast quarterly GDP with a set of 118 macroeconomic monthly data series, which compares favorably to other competing methods.
with Yongmiao Hong, Yuying Sun
PDF Slides
Abstract: With the advent of complex information systems to collect data, real-time nowcasting faces various challenges, including a more significant number of predictors, higher order of lags, unbalanced data structure, model uncertainty and complexity in bridging high-frequency information contained. To address these issues, this paper proposes a new real-time nowcasting forecast combination with dynamic factor regressions, which deletes redundant predictors and simultaneously selects optimal weights for candidate models. We show that the selected weight achieves asymptotic optimality and consistency, even when all candidate models are misspecified. The proposed estimator is consistent and asymptotically Gaussian if the true model is included in candidate models. Simulation results show that the proposed method yields lower mean square forecast errors than alternative nowcasting methods, including MIDAS in Ghysels et al. (2004), GARS in Giannone et al. (2008), and FADL-MIDAS in Andreou et al. (2013). The proposed method is applied to forecast quarterly GDP with a set of 118 macroeconomic monthly data series, which compares favorably to other competing methods.
TAKE A LOOK AT
My Updates
- All
- Conference
- Seminars
- Visiting
The 2026 ISBA World Meeting
June 28th - July 3rd
Nagoya, Japan
CES CAC Meeting
July 4th - July 5th
Chengdu, China.
Economics Seminar
September 11th
HKUST, Guangzhou, China.
Econometrics Seminar
September 25th
University of Glasgow, Glasgow, UK.
Econometrics Seminar
October 13th
University of Groningen, Groningen, Netherlands.
More about me
Contact
Office hour
By appointmentOffice
Room 3.004, Arthur Lewis Building
Department of Economics, SoSS
University of Manchester
Oxford Road, Manchester
M13 9PL
United Kingdom