A2008
Title: Censored ExponenTial smoothing to solve lost-sales demand forecasting problems
Authors: Diego Pedregal - University of Castilla-La Mancha (Spain) [presenting]
Juan R Trapero - Universidad de Castilla-La Mancha (Spain)
Enrique Holgado - Universidad de Castilla-La Mancha (Spain)
Abstract: Sales data are used as a good approximation to the true demand in many supply inventory management contexts, even though it is well-known that such an approximation is not as good as it may seem when lost sales occur. In these cases, crucial elements for inventory management, such as safety stocks or reorder points, are miscalculated. A censored general ExponenTial Smoothing algorithm is developed to estimate correctly demand when only sales are available in contexts where lost sales are known to occur. The algorithm is based on a general dynamic linear innovations state space system with censoring levels at the output. Once the model is set up, the estimation of parameters and initial conditions are estimated as usual by Maximum Likelihood. Such generality provides a solution with several advantages. First, the solution is general enough to be able to implement other methodologies straightforwardly, like time-varying regression or ARIMA models with censoring levels. Second, time-varying censoring levels may be considered, actually, the usual situation in real life, where censoring levels depend on forecasts themselves and, consequently, change over time. Finally, models with all sorts of components (like the trend, seasonal, coloured noise, and exogenous variables) can be used.
