Forecasting U.S. hurricanes 6 months in advance J. B. Elsner, 1 R. J. Murnane, 2 and T. H. Jagger 1 Received 13 January 2006; revised 11 March 2006; accepted 26 April 2006; published 31 May 2006. [1] Katrina is a grim reminder of the serious social and economic threat that hurricanes pose to the United States. Recent advances in hurricane climate science provide skillful forecasts of the U.S. hurricane threat at (or near) the start of the season. Predictions of hurricane landfalls at longer lead times (forecast horizons) for the complete hurricane season would greatly benefit risk managers and others interested in acting on these forecasts. Here we show a model that provides a 6-month forecast horizon for annual hurricane counts along the U.S. coastline during the June through November hurricane season using the North Atlantic Oscillation (NAO) and Atlantic sea-surface temperature (SST) as predictors. Forecast skill exceeds that of climatology. The long-lead skill is linked to the persistence of Atlantic SST and to teleconnections between North Atlantic sea-level pressures and precipitation variability over North America and Europe. The model is developed using Bayesian regression and therefore incorporates the full set of Atlantic hurricane data extending back to 1851. Citation: Elsner, J. B., R. J. Murnane, and T. H. Jagger (2006), Forecasting U.S. hurricanes 6 months in advance, Geophys. Res. Lett., 33, L10704, doi:10.1029/2006GL025693. 1. Introduction [2] Predictions of basin-wide Atlantic hurricane activity have been around since the middle 1980s [Gray, 1984b]. Research focusing on climate factors that influence hurri- cane frequency regionally [Lehmiller et al., 1997; Bove et al., 1998; Maloney and Hartmann, 2000; Elsner et al., 2000; Murnane et al., 2000; Jagger et al., 2001; Larson et al., 2005] is more recent. Insights into regional hurricane activity are used to help predict landfall activity [Elsner and Jagger, 2006a; Saunders and Lea, 2005; Lehmiller et al., 1997]. However, current landfall forecasts have short lead times (less than 1 month) and rely on data spanning approximately the past half century. In general, statistical models built from longer data records would be expected to perform with greater precision. However, older data tend to be less reliable and more uncertain. Here we maximize the utility of available data by combining the relatively short, high quality time series of observations with older, less precise time series using a Bayesian approach that does not require data to have uniform precision [Elsner and Bossak, 2001; Elsner and Jagger, 2004]. In doing so we offer for the first time a forecast model that can be used to predict the number of hurricane landfalls along the U.S. coastline (U.S. hurricane activity) by February 1st (4 months prior to the official start of the hurricane season and 6 months prior to the active portion of the season). The work builds on Elsner and Jagger [2006b] who demonstrate a skillful prediction model for U.S. hurricanes by July 1st. 2. Data [3] A chronological list of all hurricanes that have affect- ed the continental United States in the period 1851–2004 is available from the U.S. National Oceanic and Atmospheric Administration. The approximate length of the U.S. coast line affected by hurricanes from the Atlantic is 6000 km. We do not consider hurricanes affecting Hawaii, Puerto Rico, or the Virgin Islands. Hurricane landfall occurs when all or part of the storm’s eye wall passes over the coast or adjacent barrier islands. A hurricane can make more than one landfall as hurricane Andrew did in striking southeast Florida and Louisiana. Here we consider only whether the cyclone made landfall the continental United States at least once at hurricane intensity. Here it is assumed that the annual counts of U.S. hurricanes are certain back to 1899, but less so in the interval 1851 – 1898. Justification for this cutoff is based partly on U.S. legislation in July 1898 to create a hurricane warning system for the protection of military and merchant ships in the Caribbean that led to the establishment of a Weather Bureau forecast center at Kingston, Jamaica [Arsenault, 2005]. [4] We consider as predictors of U.S. hurricanes two variables shown previously to be related to seasonal activ- ity; Atlantic SST and the North Atlantic oscillation (NAO), represented by a sea-level pressure difference between subtropical and polar latitudes [Hurrell et al., 2001]. Atlantic SST values are based on a blend of model values and interpolated observations, which are used to compute anomalies north of the equator. The anomalies are computed by month using the climatological time period 1951 – 2000 and are available back to 1871. Units are °C. January SST values are obtained online from NOAA-CIRES Climate Diagnostics Center (CDC). The low frequency variation in linearly detrended Atlantic SST is sometimes referred to as the Atlantic Multidecadal oscillation (AMO) [Enfield et al., 2001; Goldenberg et al., 2001]. For short- hand we use the acronym ‘‘AMO’’ for Atlantic SST variation. NAO index values are calculated from sea level pressures at Gibraltar and at a station over southwest Iceland [Jones et al., 1997], and are obtained from the Climatic Research Unit. The values used here are an average over the fall and early winter months of October through January and are available back to 1851. Units are standard deviations. [5] We also consider the Southern Oscillation Index (SOI) as a predictor, but find no significant relationship GEOPHYSICAL RESEARCH LETTERS, VOL. 33, L10704, doi:10.1029/2006GL025693, 2006 1 Department of Geography, Florida State University, Tallahassee, Florida, USA. 2 Risk Prediction Initiative, Bermuda Biological Station for Research, St. George’s, Bermuda. Copyright 2006 by the American Geophysical Union. 0094-8276/06/2006GL025693$05.00 L10704 1 of 5