by Judith Curry
Climate Forecast Applications Network (CFAN)’s first seasonal forecast for Atlantic hurricanes is based on a breakthrough in understanding of the impact of global climate dynamics on Atlantic hurricane activity.
Research conducted by Senior Scientist Jim Johnstone at my company Climate Forecast Applications Network (CFAN) has identified skillful new predictors for seasonal Atlantic Accumulated Cyclone Energy (ACE) and the number of U.S. landfalling hurricanes.
CFAN’s prediction for the 2017 Atlantic hurricane season:
- ACE: 134 (average value 103 since 1982)
- # of U.S. landfalling hurricanes: 3 (average value 1.7 since 1900)
CFAN’s research team has long-standing expertise in climate dynamics and tropical meteorology research and in developing operational forecasts of tropical cyclones on timescales from 1-30 days. This seasonal forecast reflects the first time that this expertise has been integrated into a seasonal prediction of Atlantic hurricane activity.
North Atlantic hurricane activity varies substantially from year to year, including 5-fold variations in annual basin-scale Accumulated Cyclone Energy (ACE). The frequencies of U.S. and Florida landfalling hurricanes also changes sporadically, with recent decades witnessing multiple Category 3+ hurricanes in consecutive years of 2004 and 2005, two of the most active seasons of the past century, but also multi-year periods of quiescence, including an absence of major hurricane and Florida landfalls from 2006 to 2015.
There are numerous seasonal forecasts of Atlantic hurricane activity made by government labs, university scientists and private sector weather service providers. Most forecasts address # of named storms, # of hurricanes and # of major hurricanes. Several also predict Accumulated Cyclone Energy (ACE) and landfall probabilities. These forecasts are primarily based on ENSO forecasts, circulation regimes in the Atlantic, and statistical analysis of historical hurricane activity. Seasonal forecasts made prior to April show little skill, with most forecasts made in June or later showing skill relative to climatology, at least in hindcast mode. Notable recent forecast failures were the intensely active 2005 hurricane season and the low activity in 2013.
CFAN’s analysis of the climate dynamics of global circulation patterns has identified new predictors of Atlantic hurricane activity. CFAN’s analysis of the climate dynamics of Atlantic hurricanes includes consideration of four time scales:
- quasi-biennial (2-3 years), dominated by stratospheric signals
- interannual (3-7 years), dominated by ENSO
- decadal (7-16 years), dominated by Atlantic Ocean circulation patterns
- low frequency (15+ years)
Based on our predictability analysis, we have selected ACE and the total number of U.S. landfalls as predictands. The predictors used in this forecast are entirely empirical, based on data from January through May. Hindcasts and predictions using CFAN’s predictors for Atlantic ACE and U.S. landfalls are described below.
ACCUMULATED CYCLONE ENERGY (ACE)
North Atlantic ACE is an integrative metric of tropical cyclone duration and intensity based on the square of maximum sustained surface winds during Named Storms.
CFAN’s ACE forecast model was developed from historically correlated patterns of springtime anomalies over the period 1982-2016 of sea surface temperature (SST), SST tendencies, atmospheric sea-level pressure (SLP) and lower-stratospheric temperatures. To remove redundancies in the SST and SLP predictor indices, a principal component (PC) analysis was performed, yielding a single dominant mode with significant ACE correlation (r = 0.67). This index reflects higher ACE totals with low pressure and warm surface temperatures over the North Atlantic, high SST in the SW Pacific, and high SLP and cool and falling SSTs over the eastern Pacific. The unexplained residual variability in the forecasted ACE was found to have significant negative correlations with lower stratospheric temperatures above the North Atlantic (r = -0.5). This stratospheric indicator was combined with the surface-based PC Index in a linear regression model to obtain a final ACE forecast index with improved hindcast skill (r = 0.75).
Based upon statistics since 1982, the average error for CFAN’s ACE forecast is 35, compared to an error of 52 using the climatological average.
It is instructive to compare CFAN’s ACE forecasts with the CSU forecast. The average error of 35 for CFAN’s forecast compares to an average error of 30 units for CSU’s June forecast. In comparing CFAN’s ACE forecast (Figure 1) with CSU’s ACE forecast shown in Figure 2 (Figure 1 of the CSU forecast report), it is seen that the CSU forecast has greater skill prior to 1995, whereas CFAN has greater skill since 2008.
Figure 2: Comparison of observed ACE versus the CSU ACE model.
The largest outliers in both CFAN’s and CSU’s forecast are 2004/2005. Our analysis has identified stratospheric anomalies that relate to the intense hurricane activity during 2004/2005; better understanding of these mechanisms is the subject of ongoing research.
U.S. LANDFALLING HURRICANES
ACE is moderately correlated with most U.S. landfall indices, however the landfall indices display remarkably weak relationships to the same surface anomalies that heavily influence ACE. Inspection of historical hurricane tracks reveals that many tropical storms never approach the U.S. coastline, instead migrating northward and eastward over the open Atlantic. Recent research highlights notable differences between basin-scale cyclone activity and the frequency of hurricanes that reach the U.S. coastline, due in part to dynamical mechanisms that tend to increase general cyclone activity, but also inhibit U.S. landfalls. It is notable that elevated ACE totals from 1995 to 2014 overlapped with a historic 2006-2014 drought of major hurricane and Florida landfalls.
Annual hurricane landfalls for 5 U.S. regions were compiled from NOAA reports of 153 total events over the 1920-2016 period. Annual landfalls were tabulated for the Gulf Coast (TX to AL), Florida, the Mid-Atlantic (GA to MD), and New England (DE and states northward), as well as the Southeastern U.S. (excluding New England) and the Eastern U.S., including all states.
We identified several precursors in the ocean, troposphere and stratosphere that reflect circulation patterns that apparently influence hurricane tracks and U.S. landfalls. The clearest signal of this circulation regime is seen in the Arctic and the global stratospheric circulations. Using observations from the Arctic during spring, a new Arctic Index has been developed that provides the basis for CFAN’s prediction of the number of U.S. landfalling hurricanes.
The predictors for landfalls vary with the low-frequency Atlantic circulation regime. For the current regime (warm phase of the AMO), we considered the period 1995-2016. Figure 3 compares the modeled number of U.S. hurricane landfalls with observations.
For 2017, CFAN’s forecast for the number of U.S. landfalls is 2.7-3.4 (average value 3), with a mean absolute error of 1. For reference, the average number of hurricane landfalls since 1900 is 1.7. During the first half of the period, the model landfalls are generally lower than observed, whereas the modeled landfalls are slightly higher than observed since 2009. Because of the short training period (owing to the Atlantic regime change in 1995), the confidence in landfall forecast is lower than for the ACE forecast.
Figure 4 provides scatter plots of the Arctic Index versus the number of U.S. landfalls. For values of the Arctic Index > 0, there are no years since 1995 with zero U.S. landfalls. Identifying the regional location of landfalls is statistically less robust, but with the strongest signal for Florida.
COMPARISON WITH OTHER FORECASTS
How do CFAN’s forecasts for 2017 Atlantic hurricane activity compare with other forecasts?
- NOAA Climate Prediction Center forecasts a 45% chance of an above normal hurricane activity, with a predicted ACE range of 75%-155% of the median.
- The Colorado State University team under the leadership of Phil Klotzbach forecasts normal hurricane activity, with an ACE forecast of 100 (with a mean absolute error of 30) and a near average probability of U.S. landfalling hurricanes.
- Tropical Storm Risk forecasts average hurricane activity with an ACE forecast of 98 (with a mean absolute error of 48), with a 40% probability that U.S. landfalls will be above average
- CFAN forecasts an ACE value of 134, with a mean absolute error of 35, and above average U.S. landfalls (3).
CFAN’s forecast is unique in making a specific prediction of the number of U.S. landfalling hurricanes. Further, CFAN’s forecast uses new predictors that potentially provide greater prediction skill for years when there is no clear signal from ENSO (El Nino/La Nina). To date, the 2017 Atlantic hurricane seasonal forecasts have been regarded by forecast providers as highly uncertain owing to continued ambiguity regarding the forecasts for ENSO, although the latest ENSO forecast look to be flatly neutral for the rest of 2017, with slightly positive values.
FUTURE RESEARCH AND FORECAST PRODUCTS
We are continuing to conduct research on the climate dynamics of Atlantic hurricanes:
- Investigation the dynamics of stratospheric connections to Atlantic hurricane activity
- Improve understanding of the causes for the anomalous 2004/2005 hurricane activity
- Early season predictability and prediction (December and April)
- Multi-year forecasts (2-5 years)
- Predicting the next regime shift in the Atlantic
- Investigation of landfall dynamics during 1926-1971
In future, we plan to integrate the ECMWF seasonal forecasts more thoroughly into the forecast, along the lines of the paper by Kim and Webster (2010).
Further information about CFAN’s tropical forecast products can be obtained [here].
Apart from the intrinsic interest in the Atlantic seasonal hurricane forecast, CFAN’s forecast is an interesting example of the sociology of private sector research and how it differs from academic research.
If you are paying close attention, you will see that I do not provide sufficient information for this forecast to be reproduced. While reproducibility is the mantra (if not the norm) in academic research, in the private sector there are big counter incentives to giving away your ‘trade secrets’.
Underlying this forecast model is some very significant research into climate dynamics. Will it slow down academic research progress not to make the details of this research public? Maybe, but I’m not too worried about it since academic research is focused on other things.
Developing this forecast cost CFAN about $40K in salaries and overhead. CFAN has one client that is partially supporting this research and forecast product. In general, CFAN’s research is funded by the occasional government grant, client contracts, and overhead. More funding is needed for our hurricane climate dynamics research to continue and for regular seasonal hurricane forecasts to be sustainable. Sponsoring subscribers to the hurricane forecasts will receive full technical reports. We shall see how this funding model works.
So this is a very different model for climate research. With President Trump’s funding priorities and cuts, not to mention the endemic group think in academic climate research, this may turn out to be a good path to follow. Time will tell.