Next Season's Hurricanes

See allHide authors and affiliations

Science  07 Feb 2014:
Vol. 343, Issue 6171, pp. 618-619
DOI: 10.1126/science.1247759

Tropical cyclones (TCs) are a hazard to life and property (1, 2), as was tragically apparent following Super Typhoon Haiyan's landfall in the Philippines in 2013 and Hurricane/extratropical system Sandy's landfall in the New York tri-state area in 2012. Yet TCs also provide vital water, sometimes relieving drought (3). Predictions of the path and intensity of individual TCs are usually sufficiently good several days in advance that action can be taken. In contrast, predictions of seasonal TC activity months in advance must still be made more regionally relevant to produce information that can be acted on, for example, to improve storm preparedness.

Seasonal TC predictions focus on the probability of a range of outcomes integrated over broad regions, rather than the individual storms and narrower geographic foci used in 3- to 5-day weather forecasts. Whereas weather-scale TC predictions may lead to targeted actions such as evacuations, seasonal predictions are currently used to develop and price instruments to distribute risk (such as insurance). Improved skill and regional specificity of seasonal TC prediction could be useful to water resource, emergency, and energy management efforts. Furthermore, a better ability to forecast seasonal hurricanes can help build a more robust understanding of the ways in which climate controls hurricane activity, perhaps leading to increased confidence in multidecadal hurricane projections.

Basin-Wide Success

In recent years, several approaches have been developed to predict seasonal TC activity averaged over an entire basin, such as the North Atlantic or Northwest Pacific, several months before the season in question. These approaches include statistical (4) and dynamical general circulation models (57), as well as hybrid statistical-dynamical methods (810). They are used in operational seasonal TC outlooks made by meteorological agencies. Evaluated over multiple years and decades, these predictions are skillful at predicting the year-to-year changes in the total number of hurricanes, when compared to forecasts based on knowing only the long-term average or activity over the years preceding a season. The predictive skill of basin-wide activity can be seen in individual years. For example, for months prior to the 2010 season, Atlantic hurricane frequency was consistently predicted to be large (see the first figure), and 2010 was indeed the second most active hurricane season since 1970.

Learning from Failure

Even though predictions are skillful in predicting year-to-year changes in TC activity over many years, they are not perfect. A glaring example is the recent 2013 Atlantic hurricane season (see the first figure), for which nature failed to follow the almost unanimous prediction that the North Atlantic should have a normal to slightly enhanced number of hurricanes (∼6 to 9). Instead, it was one of the most anemic hurricane seasons ever recorded.

A season like 2013 is humbling. Yet only by understanding and learning from past failed predictions will the prediction community be able to successfully move forward. In disentangling the causes of the low hurricane activity of 2013, we must ask ourselves whether our prediction systems neglected something foreseeable, and then account for this in future predictions. But the predictability of the climate system has limits, and it may be that the causes of the inactivity in 2013 were inherently unpredictable. Although extreme failure is improbable in any given year, over many years its likelihood at some (unknowable) point can become substantial even in the best possible prediction system.

Seasonal North Atlantic hurricane prediction.

The very active 2010 season was successfully predicted by a range of methodologies (410), but these prediction systems generally failed for the very inactive 2013 season. Central estimates are circles; vertical bars show ranges [70% range for (8, 10); ±1σ for (47, 9)]. The legend gives the month when each prediction was issued. For (8), the predicted exceedance probabilities for the observed hurricane counts are given to the left of the vertical bar. For data, see supplementary materials.

Not just how many but where.

Observed TC tracks for 2010 (A) and 2013 (B). TC locations are shown at 6-hour intervals, with colors indicating their intensity. For data sources, see supplementary materials.

Toward Regional Information

Despite being a key first step, basin-wide predictions do not provide sufficient information for most applications, which require a capability to predict the likelihood of TC occurrence at a more regional scale. The 2010 Atlantic hurricane season presents a poignant example. As discussed above, this season was widely and successfully predicted to be very active at a basin scale. Yet a relatively modest number of hurricanes made landfall (although some of those landfalls did tragically result in fatalities). The dearth of landfalls was particularly pronounced in the U.S. coastline, where no hurricanes made landfall (see the second figure, panel A). An active basin-wide season does not necessarily translate into an active landfall season, nor does an inactive season in parts of the basin necessarily translate into an inactive season along the coasts. In 2013, the Atlantic was inactive both in basin-wide activity and landfalls (see the second figure, panel B).

Prediction efforts must be pushed beyond basin-wide TC activity toward the much more challenging goal of improving skill at the regional scale. These efforts will be helped by advances in understanding what controls the geographical distribution of TCs. For example, aspects of the regional distribution of TC activity may be connected to potentially predictable modes of climate variability and change (11, 12). Enhanced computer power has helped to increase the spatial refinement of climate models (57, 13, 14) and improved prediction methodologies (69), a crucial step for delivering skillful regional predictions. Many hazards associated with landfalling TCs (such as winds, storm surge, heavy rainfall, and flooding) vary among regions, and local information is therefore essential for predicting TC impacts.

Communicating Uncertainty

The climate system is chaotic, and all climate predictions are inherently probabilistic (making a statement of the likelihood of certain events) rather than deterministic (making specific statements about the course of the future). Relative to predictions of basin-wide activity, predictions at the regional scale are more likely both to invite action and to have larger uncertainties. It is thus crucial to develop an explicit and accurate representation of the uncertainties associated with the predictions. The potential utility of predictions would be enhanced through communication between the developers of prediction methodologies and the eventual users of the prediction products. At the least, users of predictions should demand—and be capable of using—information about past prediction performance and expected uncertainty.

Improved understanding and modeling capabilities are bringing us to the threshold of more skillful, region-specific, and explicitly probabilistic predictions of seasonal TC activity. In seizing this opportunity, scientists must acknowledge the limitations of their methods. Although we have focused on seasonal hurricane predictions, the issues we raise—prediction verification, learning from failed predictions, and correctly describing and communicating uncertainty—apply to all efforts to predict climate and its impacts.

Supplementary Materials

References and Notes

  1. Acknowledgments: Supported by NSF grant AGS-1262099 and by the NOAA Climate Program Office. We thank J. Baldwin, T. Delworth, A. Johansson, S. Kapnick, D. Lavers, and G. Saville for useful comments, and P. Klotzbach, T. LaRow, J. Schemm, F. Vitart, H. Wang, and NOAA's Climate Prediction Center for making prediction data available to us.

Navigate This Article