Adaptive Estimation of Daily Demands with Complex Calendar E ects for Freight Transportation

We address the problem of forecasting spatial activities on a daily basis that are subject to the types of multiple, complex calendar e ects that arise in many applications. Our problem is motivated by applications where we generally need to produce thousands, and frequently tens of thousands, of models, as arises in the prediction of daily origin-destination freight ows. Exponential smoothingbased models are the simplest to implement, but standard methods can handle only simple seasonal patterns. We propose a class of exponential smoothing-based methods that handle multiple calendar e ects. These methods are much easier to implement and apply than more sophisticated ARIMAbased methods. We show that our techniques actually outperform ARIMA-based methods in terms of forecast error, indicating that our simplicity does not involve any loss in accuracy.