Forecasting Stock Returns with Model Uncertainty and Parameter Instability, a paper authored by our school’s PhD student Zhang Hongwei from the Tilburg Program, under the guidance of our school’s Associate Prof. Jiang Fuwei, Prof. He Qiang, and Prof. Ben Jacobsen from the TIAS School for Business and Society, Tilburg University, was officially accepted for publication by Journal of Applied Econometrics, a world-renowned journal on applied economics. The paper puts forward a brand-new portfolio forecasting method which can significantly raise the forecasting accuracy for equity risk premium or expected stock returns, and has hence made important academic contributions to researches on empirical asset pricing, financial machine learning and quantitative investment, among others.

First of all, it comprehensively looks into the forecasting capability of multiple model averaging methods (including KS, BMA, MMA, JMA, WALS, LASSO and ENet) as well as machine learning algorithms for equity risk premium, and comes to the conclusion that the out-of-sample forecasting accuracy of most models is very low. It points out that this is because these classic models have all neglected the inherent characteristics of financial markets such as dynamism, evolution and frequent structural changes. But financial markets feature model uncertainty and parameter instability, so these models which are built on the assumption of stable data generation have all lost forecasting accuracy.

Then it comes up with a new portfolio forecasting method with model uncertainty, parameter instability and shrinkage all taken into consideration, and applies it to model averaging (such as KS and LASSO) and classic machine learning algorithms. The new method can substantially boost the out-of-sample forecasting capability of these classic model averaging and machine learning methods for equity risk premium. Moreover, it can also increase the forecasting capability for such macroeconomic indicators as GDP, economic sentiment and probability of recession. From both the statistical theory perspective of bias-variance tradeoff, and the economic theory perspective of time-varying macroeconomic risks and time-varying risk aversion, the paper explains why the new portfolio forecasting method can greatly enhance forecasting accuracy.