Skip to content

GFDL Experimental Long Lead Seasonal Hybrid Hurricane Forecast System (HyHuFS)

Key Findings

  • Understanding developed in order to explore the century-scale response of
    hurricanes to climate change has led to an improved method for forecasting year-to-year
    variations in seasonal hurricane activity.
  • Skillful forecasts of the frequency of Atlantic hurricanes are feasible
    from as early as November of the previous year.
  • The frequency of Atlantic hurricanes over a broad range of climates is well
    described using a simple statistical model with two predictors: Atlantic and
    tropical-mean sea surface temperature.
  • This website contains experimental forecasts of Atlantic hurricane frequency
    for the continued evaluation of the methodology and enhancement of our understanding of the influence of climate on hurricanes.



These web pages contain experimental forecasts of North Atlantic hurricane frequency and their continued evaluations in order to assess the performance of the system in as realistic a prediction mode as possible.

Updated 29-March-2013 14:34 EDT.

Click for Latest Experimental Forecast for 2014 (April 8 2014) – look below for earlier forecasts

NOTE: As of March 2012, the HyHuFS system moved from using v1.0 of the GFDL CM2.1 EnKF seasonal forecast system to v3.1 of the GFDL CM2.1 EnKF seasonal forecast system.

Note: These experimental hurricane forecasts are NOT an official outlook.
This is a research product on the continued verification and evaluation of an experimental forecast system. We make these experimental forecast results available in order to facilitate and motivate research and discussion on the topic of long-lead seasonal hurricane forecasts. The forecasts of basinwide hurricane activity are not forecasts of landfalling hurricane activity: years in which there are many hurricanes in the Atlantic as a whole need not be years with may hurricanes making landfall (an example of this is 2010) and vice-versa;  and since damaging landfalling hurricanes can occur even in years with little activity overall (for example, Hurricane Andrew did extreme damage to South Florida in 1992, which was an inactive year overall).


Schematic of the elements of HyHuFS

Forecast system evaluation

Below are summaries of the continued evaluation of the performance of the GFDL Hybrid (Statistical-Dynamical) Hurricane Forecast System. We will attempt to update these regularly (in the first half of each month). We seek to encourage further evaluation of this system by making these data public, and would welcome feedback. Of particular interest are novel methods to evaluate the skill of and make use of the information from probabilistic forecasts.


2012 Hurricane Season:

2013 Hurricane Season:

2014 Hurricane Season:

 Forecast system description

Recently we have developed an experimental system to make forecasts of seasonal hurricane frequency in the North Atlantic, building on our understanding of the influence of forced and internal climate changes on hurricane activity. The initial description of this methodology and its performance over the period 1982-2009 is contained in Vecchi et al. (2011, Monthly Weather Review), and indicate that it may be feasible to make skillful predictions of the seasonal frequency of hurricanes over the entire North Atlantic basin (which peaks in August-October) from as early as November of the previous year. The goal of this website is to continue the evaluation of this forecast system in order to assess its performance over as long a record as possible.

The forecasts are generated using a hybrid statistical-dynamical North Atlantic hurricane frequency prediction scheme, which involves dynamical climate models to predict the future state of the climate system and a statistical model to estimate the frequency of North Atlantic hurricanes given the predicted future climate. The statistical hurricane model was built from the response of hurricane frequency within a high-resolution global climate model (Zhao et al. 2009, 2010) to changes in sea surface temperature over a broad range of climates; the Zhao et al. dynamical model exhibits skill in describing past hurricane activity when forced with observed sea surface temperatures, and training on it allows us to build a statistical model that is more independent of observed hurricane counts and that is likely to be valid for a large range of climate conditions. The statistical model is built using the methodology of Villarini et al (2010), which uses a Poisson regression technique on a limited number of predictors, which are a parsimonious description of past hurricane activity as well as the sensitivity of hurricane frequency to a broad range of possible climates. The two indices we use are Tropical Atlantic sea surface temperature (SST) and tropical-mean SST, with tropical Atlantic SST acting as a positive predictor (when it is warm hurricane frequency tends to increase) and tropical-mean SST acting as a negative predictor (when it is warm, hurricane frequency tends to decrease). This reflects recent findings that it is the difference between the tropical Atlantic and tropical-mean SST (relative SST) that is relevant to hurricane frequency, such that when the tropical Atlantic is warmer than the rest of the tropics hurricane activity increases (and vice versa).

The dynamical forecast the two climate indices (Atlantic and Tropical-mean SST)thus far have been made using two global climate models initialized with the observed climate
record: (1) the NOAA Geophysical Fluid Dynamics Laboratory’s (GFDL) experimental seasonal to interannual (S-I) forecast system and (2) National Center for Environmental Prediction’s (NCEP) Climate Forecast System version 1 (CFS.v1). These forecast systems use computer representations of the governing physical equations of the climate system (such as conservation of energy, mass, momentum) to predict the future state of the climate system based on our best observational estimates of its present state.

The forecasts of hurricane frequency are probabilistic, reflecting the fact that hurricanes are influenced by factors that are not fully-predictable and that the influence of those factors on hurricanes is also not fully-predictable. Therefore, while for each year we present the “expected value” or average storm count one would expect to see based on all years that were equivalent to a given year (which is not necessarily the most likely value for a given year, for example the “expected value” of a single roll of a die is 3.5), we also present a series of probabilistic measures of hurricane activity. We currently focus on two probabilistic measures:

  • “Exceedence Probabilities”, which are measures of how likely it is that the hurricane counts for the season will exceed a certain threshold. For example, the exceedence probability for a single die roll of the value 1 is 83.3% (meaning that five out six die rolls, on average and over many rolls, will have a value larger than 1).
  • “Centered Confidence Intervals”, which describe the range of values – centered about the expected value, that have a certain probability of occurring in a given year. For example, the centered confidence interval at 66.6% for a single die roll is between 2 and 5, meaning that (on average, over a long enough time) one out of three rolls of a die will be lower than 2 or higher than 5.


Figure 1: Retrospective (1982-2009) performance of HyHuFS predictions of hurricane frequency initialized in January (adapted from Vecchi et al. 2011). Black line shows the observed hurricane frequency. Red line shows the predicted expected value. The pink, yellow and green shaded areas show the predicted 50%, 75% and 90% confidence range on the predictions, respectively.


Contact: Gabriel Vecchi –


Delworth, T.L., and Coauthors (2006): GFDLs CM2 global coupled climate models. Part I: Formulation and simulation characteristics. J. Climate, 19, 643?674.

Saha S, S. Nadiga, C. Thiaw, J. Wang, W. Wang, Q. Zhang, H.M. van den Dool, H.L. Pan, S. Moorthi, D. Behringer, D. Stokes, M. Pena, S. Lord, G. White, W. Ebisuzaki, P. Peng, and P. Xie (2006): The NCEP Climate Forecast System. J. Climate, 19, 3483-3517.

Vecchi, G.A., M. Zhao, H. Wang, G. Villarini, A. Rosati, A. Kumar, I.M. Held, R. Gudgel (2011): Statistical-Dynamical Predictions of Seasonal North Atlantic Hurricane Activity. /Mon. Wea. Rev, In press.

Villarini, G., G.A. Vecchi and J.A. Smith (2010): Modeling of the Dependence of Tropical Storm Counts in the North Atlantic Basin on Climate Indices. Mon. Wea. Rev., 138(7), 2681-2705, doi:10.1175/2010MWR3315.1

Zhang, S., M.J. Harrison, A. Rosati, and A.T. Wittenberg (2007): System design and evaluation of coupled ensemble data assimilation for global oceanic climate studies. Mon. Wea. Rev., 135(10), doi:10.1175/MWR3466.1.

Zhao, M., I.M. Held, S.-J. Lin, and G.A. Vecchi (2009). Simulations of global hurricane climatology, interannual variability, and response to global warming using a 50km resolution GCM. J. Climate, 22(24), 6653-6678, doi:10.1175/2009JCLI3049.1

Zhao, M., I.M. Held and G.A. Vecchi (2010): Retrospective forecasts of the hurricane season using a global atmospheric model assuming persistence of SST anomalies. Mon. Wea. Rev., 138, 3858-3868.