High Dimensional Time-varying Forecast Combination: A Unified Approach for Abrupt and Smooth Instabilities

Project: Research

View graph of relations


In recent years, scholars from the economics and finance fields have noted an increase in the number of individual forecasts. Empirical asset pricing studies also indicate that a forecast’s predictability power changes over time. Moreover, the relative importance of individual forecasts among competing predictions may vary according to economic periods. In the presence of many predictors, forecasting combinations with instabilities, which provide the opportunity to exploit a richer base of information used for time series forecasting dynamically and adaptively than conventionally, have become increasingly popular.The out-of-sample performance and evaluation of a time-varying predictive regression model directly rely on the formulation of time-varying parameters. However, the underlying form of the time-varying weights that individual forecasts exhibit on the variable to be predicted is unknown. Empirical studies often employ rolling window methods to make time-varying forecasts. However, this approach is limited in that it arbitrarily eliminates the data. In contrast, assuming specific functional forms for time variation mainly used for estimation and computation simplicity cannot recover the actual time-varying dynamics. Smoothed nonparametric estimation methods, such as local and global estimation approaches, can help reduce misspecification errors in parametric models. However, these methods are not consistent in abrupt structural breaks. Therefore, unsurprisingly, existing studies often document spurious and even controversial conclusions on the forecasting ability of the asset and bond returns of the same individual forecast when using different window sizes or different samplingperiods.Thus, it is not clear how forecasts with nonzero predictive powers can be correctly distinguished from redundant ones in the face of increasingly available forecasts. Further, related studies have not examined how the changes in their relative importance in different periods given technological developments, changes in preferences and skills, structural shifts in economic conditions, and deceptiveness to new regulatory rules can be estimated accurately. Moreover, it is not clear whether the disappearance of forecasts can be attributed to estimation inadequacy or their actual abandonment by financial markets.Therefore, it is crucial to propose a new forecast combination method that can manage many individual forecasts, whose relative importance can change over time. We develop a new approach for forecasting with many predictors, in the face of instabilities. Particularly, we allow the number of predictors to be larger than the sample size. In sharp contrast to existing forecast combination literature, our method allows for abrupt, smooth, and dual-type time variation in high-dimensional predictive regression. 


Project number9043433
Grant typeGRF
Effective start/end date1/01/23 → …