Approaches in forecasting cereal production

We assess the forecasting abilities oftransfer function model and spatial autoregression, along with the possible benefits from deseasonalization. When the deterministic seasonal components are set aside, the structural dynamics in a time series model of the remaining stochastic components are better understood, hence empirically fitted well, facilitating forecasting. However, a model that best incorporates the interactions among different agents of seasonality is still superior to that of the model that only sets aside seasonality, not making it an integral part of the model. When there is a pronounced, stable seasonality usually occurring when deterministic seasonality dominates the stochastic seasonality, deseasonalization becomes beneficial for forecasting in transfer function models.

Spatial autoregression provides an alternative modeling framework when there is a constraint on available explanatory variables. Spatial and temporal autoregressions can account for spatial externalities and temporal accumulations that explain a large part ofthe fluctuations in a time series forecasting situation. Using rice and corn production data for the Philippines, forecast errors can be reduced by at least half with the inclusion of a spatial autoregressive term in the model.