Generalized Additive Models versus Linear Regression in Generating Probabilistic MOS Forecasts of Aviation Weather Parameters

Abstract

The skill of probabilistic Model Output Statistics forecasts generated from Generalized Additive Models (GAM) is compared to that of traditional multiple linear regression techniques. Unlike linear regression, where each predictor term in the additive model is assumed to vary linearly with the predictand (unless specified otherwise by the developer), GAM is a nonparametric tool that makes use of the data to automatically estimate the appropriate functional (curvative) relationship for each predictor term. This relieves the developer from the chore of identifying and computing the correct predictor transformations and helps uncover certain nonlinearities that may have been missed. Forecast equations for each statistical technique are developed for nine regions encompassing a total of 90 stations in the northeastern United States. Three parameters (cloud amount, ceiling height, and visibility) are forecast for eight thresholds and two lead times (12 h and 24 h). The developmental dataset consists of limited-area fine-mesh numerical model output and surface observations for the period 1984?1989. Verification on 3 yr (1990?1992) of independent data indicates a clear and consistent superiority of the GAM model over linear regression, with mean square errors generally 3%?4% lower and lead time gains of 2?9 h. To some extent, GAM's additional computational burden relative to linear regression has deterred its operational implementation. However, as computer power and memory continue to increase while prices continue to fall, the time is drawing near when the use of such modern statistical techniques will be operationally feasible.