Title: Bagging with Real Data
Presenter:Fotios Petropouplos
Hosted by: Professore Kang Yangfei
Time: 2019.4.26 16:00-17:00
Location: A618
Abstract:
The existing portfolio of models within a forecasting software is, very often, not able to capture the true data generating process (DGP). This had led many researchers to combine the forecasts from two or more forecasting models and, on average, achieve better accuracy. One approach to forecasting combination is bootstrapping and aggregation, or bagging. In this approach, the remainder of the data from a decomposition method is bootstrapped and then used to create new instances of the original series. Each new series is forecasted separately and the forecasts are then combined. In this talk, we present a new approach where instead of bootstrapping the remainder to create a new series, we use a large database to find similar time series. These similar series are either forecasted using familiar models, explicitly assuming a DGP, or their “future” values are directly taken as forecasts in which case no DGP is assumed. Our approach is tested on real data and shows promising improvements over standard benchmarks, especially when historical information is limited.
About the presenter:
Fotios Petropoulos is Associate Professor at the University of Bath, UK, and Associate Editor for theInternational Journal of Forecasting. He is interested in research on time series forecasting, judgmental approaches for forecasting, statistical and judgmental model selection and integrated business forecasting processes. Fotios's research so far has focused on the improvement of forecasting processes and more specifically around two streams. First, he has examined how additional information can be extracted from time series data through time transformation (temporal aggregation) and the use of hierarchies. Second, he has investigated interactions between forecasting and management judgment.
