J15b.3 Robust and Interpretable Statistical Models for Predicting the Intensification of Tropical Cyclones

Based on the tropical cyclone (TC) forecasting literature, hurricane intensity prediction is one of the most challenging tasks. To date, the problem has been addressed through statistical models (SHIFOPR, ST5D), statistical-dynamical models (SHIPS) and primitive equation numerical models (GFDL), predicting the intensity changes for up to five days. For the first two kinds of models, both multiple linear regression (MLR) and non-linear regression in the form of neural networks (NNs), have been applied with promising results (DeMaria et al. 2005, Castro 2004, Knaff et al. 2004, Baik and Hwang 2000, Baik and Hwang 1998). On the other hand, the procedures for feature selection and for reporting the predictive performance of the derived models have not been investigated to a great extent, in the sense that (1) they widely vary, so comparisons between models are made in an ad-hoc basis; (2) the derived models have an inherent selection bias, i.e. allowance to peek in the test set during feature selection, prohibiting good generalization behavior (Ambroise and McLachlan 2002); and (3) they are unstable in terms of performance and understanding (Guyon and Elisseeff 2003). For example, it is often the case that at certain seasons the models perform extremely well and in others quite unsatisfactory, while there is a constant update in the set of features used, lowering the interpretability of the models. Having the above in mind, the goal of this paper is twofold: (a) to build robust models; and (b) build models that are explicitly or implicitly interpretable, delivering additional knowledge about the problem. Robustness in this context can be defined as a property of a model that is performing efficiently, is able to generalize well and is parsimonious – a characteristic of models that generalize well – based on what Ockham's razor principle implies: complexity (in our case extra features) must pay for itself by giving a significant improvement in the error rate during the training procedure (Cristianini and Shawe-Taylor 2000). This principle is quantified in section 3. Recently developed rule based regression schemes are also a focus of the work presented here. They are applied to the dataset in order to identify more elaborate structure behind the intensity predictions. MLR and NNs fail to provide the human expert with