Hurricane trend detection

Abstract

Because a change in the frequency (number/year) of hurricanes could be a result of climate change, we analyzed the historical record of Atlantic basin and US landfalling hurricanes, as well as US continental accumulated cyclone energy to evaluate issues related to trend detection. Hurricane and major hurricane landfall counts exhibited no significant overall trend over 167 years of available data, nor did accumulated cyclone energy over the continental USA over 119 years of available data, although shorter-term trends were evident in all three datasets. Given the χ2 distribution evinced by hurricane and major hurricane counts, we generated synthetic series to test the effect of segment length, demonstrating that shorter series were increasingly likely to exhibit spurious trends. Compared to synthetic data with the same mean, the historical all-storm data were more likely to exhibit short-term trends, providing some evidence for long-term persistence at timescales below 10 years. Because this might be due to known climate modes, we examined the relationship between the Atlantic multidecadal oscillation (AMO) and hurricane frequency in light of these short-term excursions. We found that while ratios of hurricane counts with AMO phase matched expectations, statistical tests were less clear due to noise. Over a period of 167 years, we found that an upward trend of roughly 0.7/century is sufficient to be detectable with 80% confidence over the range from 1 to 21 storms/year. Storm energy data 1900–2018 over land were also analyzed. The trend was again zero. The pattern of spurious trends for short segments was again found. Results for AMO periods were similar to count data. Atlantic basin all storms and major storms (1950–2018) did not exhibit any trend over the whole period or after 1990. Major storms 1950–1989 exhibited a significant downward trend. All-storm basin scale storms exhibited short-term trends matching those expected from a Poisson process. A new test for Poisson series was developed based on the 95% distribution of slopes for simulated data across a range of series lengths. Because short data series are inherently likely to yield spurious trends, care is needed when interpreting hurricane trend data.