Predicting spring tornado activity in the Central Great Plains by March 1st

Here we illustrate a statistical model for predicting tornado activity in the central Plains by March 1st.  The model predicts the number of tornado reports during April--June using February sea-surface temperature (SST) data from the Gulf of Alaska (GAK) and the western Caribbean region (WCA).  The model uses a Bayesian formulation where the likelihood on the counts is a negative binomial distribution and where the non-stationarity in tornado reporting is included as a trend term plus first-order autocorrelation.  Posterior densities for the model parameters are generated using the method of integrated nested Laplacian approximation (INLA).  The model yields a 51% increase in the number of tornado reports per degree C increase in SST over the WCA and a 15% decrease in the number of reports per degree C increase in SST over the GAK.  These significant relationships are broadly consistent with a physical understanding of large scale atmospheric patterns conducive to severe convective storms across the Great Plains.  The SST covariates explain 11% of the out-of-sample variability in observed F1--F5 tornado reports.  The paper demonstrates the utility of INLA for fitting Bayesian models to tornado climate data.  The research was conducted in the Department of Geography at Florida State University in collaboration with Holly Widen.  It will be published later this year in the American Meteorological Society's Monthly Weather Review.  The code is available from http://rpubs.com/jelsner/4745.

Five year model of Atlantic hurricanes

Hurricanes cause drastic social problems as well as generate huge economic losses. A reliable forecast of the level of hurricane activity covering the next several seasons has the potential to mitigate against such losses through improvements in preparedness and insurance mechanisms. We develop a statistical model to predict North Atlantic hurricane activity out to five years. The algorithm has two components, a time series model to forecast average hurricane-season Atlantic sea surface temperature (SST), and a regression model to forecast the hurricane rate given the predicted SST value. The algorithm uses Monte Carlo sampling to generate distributions for the predicted SST and model coefficients. For a given forecast year, a predicted hurricane count is conditional on a sampled predicted value of Atlantic SST. Thus forecasts are samples of hurricane counts for each future year. Model skill is evaluated over the period (1997--2005) and compared against climatology, persistence, and other seasonal forecasts issued during this time period. Results indicate that the algorithm will likely improve on earlier efforts and perhaps carry enough skill to be useful in the long-term management of hurricane risk. Read more.