Introduction

Nearly a decade ago, Brooks et al.1 revealed a dichotomous trend in tornado activity across the United States, specifically, a decrease in the number of tornado days per year yet an increase in the number of tornado outbreak days per year. In essence, ref. 1 discovered an apparent tendency for tornadoes to cluster in greater numbers. This compelling trend applies to tornado reports that were aggregated over individual years and over the entirety of the contiguous United States (CONUS), thus fundamentally motivating the following two inter-related questions: When, during the calendar year, and where, geographically, have such changes in tornado days and tornado outbreak days occurred? For reference, the calendar-year occurrence matters because of the varying levels of weather-emergency preparedness during the months of boreal spring versus boreal winter, and thus because of the resultant human impact during these months2. Similarly, geographical location matters because of the varying population densities, socio-economic distributions, and housing conditions across the tornado-prone regions of the southern Great Plains versus the Southeast U.S., and thus because of the differing tornado vulnerability in these regions3,4. Our granular investigation herein of the seasonality and regionality of tornado-day and tornado-outbreak-day trends builds upon related work reported by refs. 5,6.

Another question that follows from the analysis in ref. 1 is whether the trends have continued over the most recent decade. This is relevant to ask in the context of anthropogenic climate change (ACC), especially given projections of robust future increases in the frequency of the basic meteorological conditions supportive of intense, potentially tornadic thunderstorms7,8,9,10. It is also plausible, on the other hand, that the time series of tornado activity is modulated by a multi-decadal cycle driven by internal modes of climate variability and teleconnections1,11,12,13,14. Herein we use the NOAA storm prediction center (SPC) database of tornado reports across the CONUS from 1960 to 2022 to extend the analysis of ref. 1 by 9 years and address this and the two other aforementioned questions.

Results

For consistency, we essentially followed the methodology of ref. 1, and defined a tornado day as any calendar day with at least one (E)F-1+ tornado, and a tornado outbreak day as any day with more than 30 (E)F-1+ tornadoes1. Over the 63-year period of our analysis, the linear trend of such tornado days is negative, with a statistically significant decrease of 1.03 days per year; the linear trend of such tornado outbreak days is positive, with a statistically significant increase of 0.04 days per year (Fig. 1a). However, an examination of the time series of these tornado activity metrics reveals reversals of these trends during the most recent decade (open circles in Fig. 1b). The linear trend of the number of tornado days since 2000 is weakly positive (0.11) and statistically insignificant (p-value of 0.73); this is due in part to the relatively high level of tornado activity in 2017. The linear trend of the number of outbreak days since 2000 is weakly positive (0.002) and insignificant (p-value of 0.97); this is due in part to the lack of tornado outbreaks in the late 2010s1. Implied here is the need for caution in interpreting Fig. 1b, especially given the sample size associated with the use of a 30-tornado threshold for outbreaks and the large year-to-year fluctuations.

Fig. 1: Annual observations of tornado days and outbreaks.
figure 1

a Linear trends (dashed lines) of tornado days and of tornado outbreak days with >10, >20, and >30 (E)F-1+ tornadoes with corresponding slopes/p-values. b The number of days per year with at least one (E)F-1+ tornado (small circles) and >30 (E)F-1+ tornadoes (small triangles)1. Large circles and solid lines are decadal means, centered on the decade (e.g., 1965–1975 for 1970). The open circles represent the final decadal mean taken in 2017 (i.e., 2012–2022).

Full size image

Indeed, there is no universally accepted definition of a tornado outbreak15. The issue with the 30-tornado threshold is that it yields a small number of annual outbreak days (the 63-year mean is 1.3 days per year), thus making meaningful statistical analysis challenging. We addressed this by re-computing the trends of tornado outbreak days using thresholds of >10, and >20 (E)F-1+ tornadoes (Fig. 1a). The linear trends using both thresholds are positive and statistically significant1,15,16. Of note is that a trend reversal in the most recent decade is also apparent in tornado outbreak days with a 20-tornado threshold. Although tornado outbreak days with a 10-tornado threshold do not exhibit a trend reversal in the most recent decade, the time series exhibits strong multi-decadal variability.

To understand possible seasonal dependencies of the long-term changes in annual tornado days, we disaggregated the time series into individual months. This approach follows ref. 17 and offers more temporal granularity than other studies6,15,18,19. Figure 2 reveals statistically significant negative trends in tornado days during the months of March through September. These tornado-day losses are particularly dramatic in June, July, and August, as was also shown by ref. 17. During the months of October through February, trends in tornado days are insignificant. These collective changes have implications on the tornado-day seasonality, as specifically represented by a 3-week shift in the peak probability of a U.S. tornado day, from 14 June during the 1960–1980 period to 24 May during the period 2000–2021 (Fig. S1). More modest shifts based on tornado reports have been identified over the southern and central Great Plains20.

Fig. 2: Annual observations of tornado days by month.
figure 2

Annual tornado days by month (blue circles) with corresponding linear trends (red lines) and p-values.

Full size image

To further explore these changes, we computed tornado-day trends during warm- versus cool-season as a function of U.S. states in tornado-prone regions (Fig. 3, S2c). Here, the warm season is defined as April-July, and the cool season as November-February, as supported by Fig. S3 and ref. 2. Of note is the statistically significant mean decrease of 3.31 tornado days per decade experienced in Texas during the warm season, which is the largest decrease of any analyzed U.S. state. Given the large geographical area of Texas and its contribution to the U.S. tornado climatology, such a decrease would, in turn, represent a large contribution to the negative warm-season trends revealed in Fig. 221. Elsewhere across the U.S., decreases in warm-season tornado days per decade have also occurred since 1960, especially in the states comprising the southern and northern Great Plains. Decreases in warm-season EF1+ tornado reports per decade have likewise occurred in the southern and northern Great Plains (Fig. S2a), implying that the warm-season tornado activity in U.S. states within the Great Plains may have contributed little to the positive trend in tornado outbreak days revealed in Fig. 1. In contrast, increases in warm-season EF1+ tornado reports per decade across the states comprising the Southeast (Fig. S2b), as revealed by limiting the analysis to the past three decades, imply that warm-season tornado activity in the Southeast U.S. may have contributed much to the positive trend in outbreak days. A regionalized, seasonalized analysis of tornado outbreak days supports both statements (Fig. S4).

Fig. 3: Warm and cool season tornado days by state since 1960.
figure 3

Trends of the annual warm (red) and cool (blue) season tornado day trends indicating the number of tornado days lost or gained per decade since 1960. Bold values represent a statistically significant change. Green and yellow shading indicate the Southern Great Plains and Southeast regions, respectively.

Full size image

Despite the lack of a trend in cool-season tornado days across the CONUS (Figs. 2, 3, S2c), Southeast states have generally experienced increases in tornado days (Fig. 3) and in tornado reports per decade (Fig. S2a), which is consistent with ref. 19, and the trends are amplified if we again limit the analysis to the past three decades (Fig. S2b, c)21. It is noteworthy that no state considered here had a statistically significant decrease in tornado days during the cool season. As supported by Fig. S4, we speculate that cool-season tornado activity in the Southeast is also a key contributor to the increasing trend in U.S. tornado outbreak days. The increase in Southeast tornado activity during both seasons is of consequence especially because of the high vulnerability to tornadoes in this region4,22.

The preceding analyses have provided insight into the “when” and “where” questions raised regarding the aggregated trends of tornado days (Fig. 1); analogous insight regarding tornado reports has been provided by ref. 2. Note that the dichotomy in the tornado-day versus tornado-outbreak trends was used to advance the narrative that in the United States, more tornadoes appear to be occurring on fewer—or at least comparable—numbers of days1. In Fig. 3, we attempt to further address the “when” and “where” questions underlying this narrative. As guided by Fig. S1 and ref. 19, we selected two regional domains, namely, the Southern Great Plains and Southeast, for a warm- and cool-season analysis of the frequency of days wherein the number of (EF1+) tornadoes exceeded different thresholds over three periods, 1960–1979 (period 1), 1980–1999 (period 2), and 2000–2022 (period 3)5.

For the Southeast region, the frequency of tornado days in the warm season (Fig. 4a, first bar) exhibits a statistically significant decrease from period 1 to 2, and from period 2 to 3, and thus has a negative overall trend (Fig. S2). In contrast, the frequencies of days with many (i.e., >10, >15, >20, >25) tornadoes exhibit statistically significant increases from period 2 to 3 (Fig. 4a) and thus positive overall trends which supports past work23. We conclude here that warm-season tornado activity in the Southeast U.S. contributes to the “fewer days, more tornadoes” tendency over the past 60 years1,24. Cool-season tornado activity in the Southeast U.S. essentially contributes to this tendency as well: Although the frequency of cool-season tornado days does not show a significant trend over the entire period (Fig. 4a and S2), the frequencies of days with many tornadoes exhibit statistically significant increases2. These results in the Southeast are consistent with ref. 24 and support the regional shifts in tornado activity that have been documented18,23,25.

Fig. 4: Regional tornado days with increasing thresholds by season.
figure 4

Frequency of days wherein the number of EF1+ tornadoes exceed the indicated thresholds for warm season (red) versus cool season (blue) months over the (a) Southeast domain and (b) Southern Great Plains domain (see Fig. 3). Positive (negative) signs indicate a significant increase (decrease) in tornado days per threshold, performed using a two-sample t-test, with greater than the indicated number of tornadoes from period 1 to period 2; from period 2 to period 3. The columns show the number of tornado days, normalized for the number of years in each period, on a logarithmic y-axis.

Full size image

For the Southern Great Plains region, the frequency of warm-season tornado days (Fig. 4b, first bar) exhibits statistically significant decreases from period 1 to 2 and period 2 to 3, and in fact has the largest overall decrease per region or season (Fig. S2). The frequencies of days with many tornadoes exhibit increases in the >15 threshold category from period 2 to 3, but no significant changes in others (Fig. 4b). As further supported by the negative overall trend in warm-season, Southern Great Plains tornado reports since 1960 (Fig. S2), we conclude here that warm-season tornado activity in the Southern Great Plains U.S. can be better characterized by a tendency for “fewer days, fewer tornadoes”. In light of the lack of significant increases in days with many tornadoes in the cool season, and of the negative overall trend in cool-season, Southern Great Plains tornado reports since 1960, we conclude that cool-season tornado activity in the Southern Great Plains U.S. can likewise be characterized by a “fewer days, fewer tornadoes” tendency. These tendencies support similar results found by refs. 5,6.

A subsidiary question following ref. 1 regards whether the trends in the regionally and seasonally disaggregated tornado reports are accompanied by trends in interannual variability, as was documented over the CONUS1. Following the methodology from ref. 26, interannual variability is assessed here using the annual differences of E(F)-1+ tornado reports over the CONUS, as well as over the Southeast and Southern Great Plains regions during the warm and cool seasons. The time series of annual differences in the regionalized, seasonalized tornado reports show generally larger variability during the warm seasons for both regions (Fig. S5). Changes in variability over time can be quantified through a form of volatility computed using a 15-year rolling standard deviation of the annual differences26. The 15-year volatility over the CONUS shows a sharp increase in the 2000s1. The large volatility in the recent two decades is mainly attributed to the warm season tornado activity in the Southeast. However, the year 2011, which was an anomalously active year, makes a substantial contribution to the high volatility in the past two decades. The removal of 2011 leads to a more gradual increase in volatility, and the resultant time series of volatility over the CONUS shows a multi-decadal cycle, with a decrease in the 1970s and an increase in the past two decades. The sensitivity of the data to one single year calls for caution regarding the role of internal climate variability and anthropogenic forcing.

Discussion

Our temporal and geospatial disaggregation of tornado reports in the United States from 1960 to 2022 has revealed herein that the long-term decreases in annual tornado days shown by ref. 1 have occurred during boreal spring and summer, and most dramatically over the months of June through August1. Such warm-season tornado-day losses are prominent within Texas and other states comprising the Southern Great Plains: Within this region, the tendency of tornado activity is that of “fewer days, fewer tornadoes”. These temporal changes have affected the tornado seasonality, so that the calendar day of peak tornado probability has shifted from the middle of June to the last quarter of May.

In contrast, warm- and cool-season tornado activity in the Southeast U.S. have contributed to a “fewer days, more tornadoes” tendency over the past 60 years. Indeed, this is due to statistically significant increases in the frequencies of days with many tornadoes, i.e., outbreaks, regardless of whether we define a tornado outbreak using >10-, >20-, and >30-tornado thresholds. Since outbreaks tend to include tornadoes of high-end intensity and have a significant impact on life and property, especially in vulnerable communities in the Southeast, future efforts to understand the range of their predictability should be prioritized.

There are indications that the dichotomous linear trends in tornado days and tornado outbreaks in the United States have relaxed over the most recent decade. It is an open question as to whether this has been driven primarily by internal modes of climate variability, or by a convolution of anthropogenic climate change and such internal modes. The ACC can affect tornado activity via the large-scale circulation changes, such as the jet stream and its variability associated with the north-south temperature gradient17,27. On the other hand, the impacts of certain multi-decadal modes13,14 may help explain the potential reversal of the trends in the recent decade.

A limitation to this study, as in all studies related to tornado occurrence, is a reliance on a report-based dataset, which has known issues and biases and a relatively short record length1,23,28. Analyses of meteorological proxies could be used to reinforce the conclusions reached herein8,10,29.

Methods

Data

We used the NOAA Storm Prediction Center severe weather database, which includes georeferenced tornado report information with time, date and (E)F rating. The limitations and issues with this dataset are well-known and discussed widely28. Our primary analysis of (calendar) tornado days helps mitigate some of these limitations and issues28. Tornadoes rated as (E)F-0 were not analyzed owing to the inherent uncertainty associated with their damage rating and their minimal impact on communities1. There is a risk that actual strong tornadoes were ranked (E)F-0 because they caused insignificant damage, but this number should be small and not affect the results.