报告题目：Tensor Forecast Combinations
嘉宾简介：Tae-Hwy Lee is Professor of Economics at University of California Riverside. His research and teaching area is econometrics, with research focus on time series, forecasting, and financial econometrics. He has published over 80 papers on nonstationary time series, specification testing, forecasting, volatility models, quantile models, factor models, nonparametric methods, shrinkage and model averaging methods, causal inference, panel data models, entropy and information theoretic methods, financial risk models, machine learning methods, and high dimensional models. He has received several awards including the NSF/ASA/BLS Senior Research Fellowship, the Korea Sanhak Foundation Award, the Bank of Korea Research Award, and the Econometric Theory Tjalling C. Koopmans Prize. He is a Senior Fellow of the Rimini Center for Economic Analysis (RCEA). He received a Ph.D. in economics in 1990 from University of California San Diego under the supervision of Sir Clive W.J. Granger (Nobel Laureate in Economic Science, 2003). His studies were also guided and influenced by Halbert White and Robert Engle (Nobel Laureate in Economic Science, 2003). He received his undergraduate degree in economics from Seoul National University in Korea.
报告摘要：We propose Tensor Forecast Combination, a new framework for combining forecasts that are naturally organized as tensors. Our method primarily exploits the underlying multiway structure of the data, instead of directly estimating the weights to each individual forecast, we estimate the overall weighting vector through mode-specific weighting vectors, thereby reducing the effective dimension of the optimization problem. We use an Alternating Minimization (AM) algorithm to iteratively estimate the weighting vectors and establish their asymptotic properties. The theoretical results show that, with consistent initial values, the tensor forecast combination estimator can achieve a faster convergence rate than the classical vector forecast combination estimator. Monte Carlo simulations support the theoretical findings. When the forecast pool exhibits a tensor structure or deviates modestly from it, tensor forecast combination yields substantial improvements in both estimation accuracy and forecasting performance. Two empirical applications to forecasting stock market volatility using Fama-French characteristic-sorted tensor data and to forecasting real GDP growth using multinational macroeconomic variables further demonstrate that tensor forecast combination delivers lower MSFE than vector forecast combination.
报告时间：2026年10月21日(周三)，16:30-18:00
线下地点：厦门大学经济楼C108
腾讯会议：245 242 055
