CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Multistep ahead forecasts: Parsimonious approach using varying coefficient models Authors:  Julija Tastu - Maersk Line (Denmark) [presenting]
Abstract: There are two common ways for producing multistep ahead predictions: direct and iterative. The direct is based on using a one-step ahead model multiple times: the prediction from the prior time stamp is used as input when generating the following step ahead forecast. The iterative approach is based on building a separate model for each prediction horizon. This recursive strategy is more efficient if the model is correctly specified which is hard to achieve in practice. The direct approach is generally more robust to model misspecifications and allows to incorporate horizon specific predictors. Computational challenges arise with the direct strategy when the number of horizons grow large. Moreover, treating each prediction horizon independently is not an efficient usage of data, as one would naturally expect the influence of predictors to be similar for the neighbouring prediction horizons. Addressing the above we propose to formulate the problem using varying coefficient models. Model coefficients are then described as smooth functions of prediction horizon, allowing to use information from the neighbouring steps when inferring for the specific horizon of interest. Multistep ahead forecasts can then be obtained by estimating parameters at a smaller number of fitting points. The model is applied operationally when forecasting shipping demand from 1 to 91 days ahead. The results are compared to the direct approach.