José D. Bermúdez, Ana Corberán-Vallet, José V. Segura, Enriqueta Vercher




We introduce an extension of exponential smoothing to deal with covariates and double seasonality that could easily be adapted to more than two seasonal cycles. Assuming additive effects and a stochastic component given by independent, homoscedastic normal errors, the exponential smoothing model can be expressed as an equivalent linear dynamic model with a very peculiar structure of the covariance matrix. The covariance matrix is a function of the unknown smoothing parameters only, while the mean vector only depends on the unknown initial conditions. These facts allow for a simplification of the statistical analysis of the model. Following the Bayesian paradigm, we obtain the joint posterior distribution of all the unknowns. Only the marginal posterior of the smoothing parameters is analytically intractable and has to be approached using simulation techniques. The conditional distribution of initial conditions giving the smoothing parameters is well-known and it can be integrated out exactly in order to compute the predictive distribution. Finally, we propose to integrate out the smoothing parameters using Monte-Carlo techniques, obtaining an estimate of the predictive distribution as well as their main characteristics: point forecasts and prediction intervals. We apply this methodology to electricity demand forecasts using two real data sets.


Lecture Notes in Management Science (2011) Vol. 3: 33-38

3rd International Conference on Applied Operational Research, Proceedings

© Tadbir Operational Research Group Ltd. All rights reserved.



ISSN 2008-0050 (Print)

ISSN 1927-0097 (Online)




·         Introduction

·         Bayesian Analysis Of The Innovations State Space Model

·         Applications To Electricity Demand Forecasting

·         References


Full Text PDF