El Niño–Southern Oscillation (ENSO) modulates severe thunderstorm activity in the US, with increased activity expected during La Niña conditions. There is also evidence that severe thunderstorm activity is influenced by the Arctic Oscillation (AO), with the positive phase being associated with enhanced activity. The combined ENSO–AO impact is relevant for situations such as in early 2021, when persistent, strong positive and negative AO events occurred during La Niña conditions. Here we examine the relation of a spatially resolved tornado environment index (TEI) with ENSO and the AO in climate model forecasts of February, March, and April conditions over North America. Bivariate composites on Niño 3.4 and AO indices show that TEI predictability is high (strong signals and probability shifts) when the ENSO and AO signals reinforce each other and low when they cancel each other. The largest increase in the expected value and variance of TEI occurs when Niño 3.4 is negative, and the AO is positive. Signal-to-noise ratios are higher during El Niño–negative AO than during La Niña–positive AO, but probability shifts are comparable.

El Niño–Southern Oscillation (ENSO) modulates severe thunderstorm activity (tornadoes, large hail,
and damaging straight-line winds) in the US, with increased activity
expected during La Niña conditions in winter and spring

Climate predictability studies can provide an indication of how much
of the observed variability is explained by ENSO and other predictable
signals and what skill is to be expected from forecasts. A general
challenge in estimating climate predictability is that the predictable
variability (signal) is usually modest compared to the unpredictable
variability (noise). Predictability can be estimated using
observations or physics-based models. An advantage of model-based
predictability studies is that sample sizes can be substantially
larger than the observational record, and both signal and noise can be
better estimated

The difference between the expected ENSO response and what occurs in a
given year (e.g., in early 2021) can be attributed to atmospheric
noise. However, such atmospheric noise may include predictable
variability that is independent of ENSO. For instance, ENSO-based
predictability of California winter precipitation is relatively low

The AO is a dominant mode of hemispheric variability which influences
North American near-surface temperature and precipitation, especially
during the cold season

A limitation of previous studies is that they have not
considered spatially resolved, sub-annual, combined impacts of ENSO
and AO on severe thunderstorm activity. Pooling data across the
US or multi-state regions can mix different climate signal
responses, which may dilute or obscure signals and their spatial
dependence. Subseasonal analysis is also preferable because of the
strong annual cycle in US severe thunderstorm activity and because
the persistence of the AO as measured by its autocorrelation function
is limited to less than 30 d

Here, we examine ENSO and AO signals in monthly climate model
forecasts. The number of monthly samples in the climate model data is
larger than the observational record by more than a factor of 200
because of multiple forecast initializations and ensemble members.
Since the climate model does not resolve thunderstorms, we investigate
the predictability of a spatially resolved tornado environment index
(TEI). TEI is known to capture some aspects of the observed tornado
climatology and variability when it is computed from reanalysis or
climate model forecast data

The Climate Forecast System, version 2

Niño 3.4 and Z1000 anomalies were computed with respect to a
forecast climatology that is a function of target month and lead time,
where lead time is defined as the number of days from the
initialization day to the beginning of the target month and ranges
from 1 to 276 d (

The tornado environment index (TEI) was computed from CFSv2 output on
a

We applied EOF analysis to CFSv2 monthly forecasts of hemispheric
Z1000 poleward from 20

For composites, positive and negative ENSO and AO conditions were
defined as occurring when index amplitudes exceeded 0.76 times the
monthly standard deviation of the index. We used a lower threshold
than the 1-standard-deviation threshold in

We measured the predictability of TEI during univariate (e.g., El
Niño) and bivariate (e.g., El Niño and positive AO) composite
conditions using skill scores that were computed under the perfect
model assumption. No observational data were used. We measured the
predictability of deterministic forecasts using the mean squared error
skill score (MSESS). The perfect model MSESS is

We computed the empirical (no fitting) cumulative probability
distribution function of the area-weighted sum of TEI over land points
east of 110

To assess the statistical significance of composite, correlation, and
probability maps, we followed the procedure of

For plotting composite, MSESS, and probability maps, we masked locations where the values were statistically insignificant or where the absolute value of the TEI composite anomaly was less than 0.05. In addition, we masked MSESS values less than 0.05 and probability values that were less than 5 percentage points away from 50 %. Our use of thresholds in addition to statistical significance reflects the fact that with large sample sizes, nearly all results are statistically significant.

ENSO and AO composites of TEI anomalies show signals that are centered
over Louisiana and Arkansas in February and that shift northward in
March and April (Fig.

Rows 1–4: univariate Niño 3.4 and AO composites of
February, March, and April TEI anomalies. TEI units are number of
tornado reports per 1

Signal amplitudes are highest in March and lowest in April. Signals in
May are weaker still (not shown). Regression and correlation maps show
the same subseasonal variation in the strength of the relation of TEI
with the Niño 3.4 and AO indices (Figs. S4 and
S5). Correlation maps show additional continental-scale structure in the west and north, where there are sizable
correlations but where the TEI variability is too small to appear in
the composite or regression patterns. The strongest TEI correlations
are of the order of 0.3–0.4, which is highly statistically significant in
the model data here but would be less so in 40 years of data for which the
5 % significance threshold would be about 0.32. The ENSO and AO
spatial patterns and amplitude are similar in February and March. In
April the AO pattern is shifted further northward than the ENSO
one. Correlation maps indicate that both TEI ingredients contribute to
the April differences between ENSO and AO patterns (Figs. S6 and S7). Overall, the ENSO signal
is slightly stronger than the AO one (Fig.

To address the question of how ENSO and AO might modulate the total
number of tornadoes, we examined the distribution of TEI summed over
land points east of 110

We computed bivariate composites of TEI anomalies conditional on the
simultaneous values of the Niño 3.4 and AO indices to investigate
the constructive and destructive interference between the ENSO and AO
signals. TEI signals are strong when the ENSO and AO signals
reinforce each other (interference is constructive), which is the case
for opposite-signed indices, namely, La Niña–AO

Rows 1–4: bivariate Niño 3.4–AO composites of February,
March, and April TEI anomalies. TEI units are number of tornado
reports per 1

Summed TEI return levels deviate from their climatological (All)
values only when the ENSO and AO signals reinforce each other (bottom
row Fig.

Perfect model MSESS values can be interpreted as squared anomaly
correlation values and are low when the ENSO and AO signals cancel
(second and third rows of Fig.

Perfect model mean squared error skill score (MSESS) for
February, March, and April bivariate Niño 3.4–AO composites of
TEI. MSESS values are masked as in Fig.

For probabilistic forecasts during each of the four bivariate
conditions, we considered the probability of TEI exceeding its
climatological median. The expected Brier skill score and log skill
score are even increasing functions of the forecast probability
shift away from its climatological value of 50 % – larger probability
shifts mean larger expected skill scores. Therefore, we only show the
probability shifts for the bivariate composites (Fig.

Probabilities of February, March, and April TEI exceeding its
climatological median value for bivariate Niño 3.4–AO
composites. The color bar is centered on the climatological value of
50 %. Statistically insignificant values, locations with composite
amplitudes less than 0.05, and shifts away from 50 % that are less
than 5 percentage points are masked. All statistically insignificant
probability shifts away from 50 % are less than

Both predictability measures show some regions west of 100

Reports of US tornadoes appear to have diverged from the
enhanced activity that would be expected during the La Niña
conditions of early 2021, a period when notable monthly Arctic
Oscillation (AO) anomalies also occurred. To investigate the question
of how ENSO and the AO jointly modulate North America severe
thunderstorm activity, we computed a tornado environment index (TEI)
in 41 years of climate model forecasts for target months in the range
February–April. Because the forecasts have many initializations and
ensemble members, the sample size is large enough to compute robust
bivariate composites based on simultaneous values of the Niño 3.4
and AO indices. Because lead times extend up to about 9 months when
forecasts are nearly independent of the verifying observations, model
results are less closely tied to the observational record of the
particular weather events that occurred. Our main findings are as follows:

ENSO and AO teleconnections in TEI have similar patterns and amplitudes over North America, with the AO index being overall positively correlated with TEI.

TEI predictability is high (strong anomalies and probability shifts) when the ENSO and AO signals reinforce each other (opposite-signed Niño 3.4 and AO indices).

When the ENSO and AO signals interfere destructively (same-signed Niño 3.4 and AO indices), the signals cancel, and TEI predictability is small.

We computed the predictability of TEI by target month conditional on the simultaneous phases of ENSO and the AO. Predictability was measured using skill scores that were computed under the perfect model assumption. The mean squared error skill score (MSESS) is a skill score for deterministic forecasts, and the perfect model MSESS depends only on the signal-to-noise ratio. To a first approximation, MSESS reflects the TEI signal amplitude and is small (little predictability) when the ENSO and AO signals cancel and is large (high predictability) when they reinforce each other. MSESS is highest in March and lowest in April. Comparing the two constructively phased situations, MSESS is higher during inactive phases (positive Niño 3.4 and negative AO indices) than during active phases (negative Niño 3.4 and positive AO indices). The reason for this difference is that the noise variance is smaller during inactive phases, and consequently the signal-to-noise-ratio is larger.

On the other hand, the perfect model Brier and log skill scores depend
only on the size of the probability shifts, which are nearly the same
for active and inactive constructively phased composites. This
difference between predictability as measured by MSESS and
predictability as measured by probability shift is perhaps unexpected
because previous studies have noted a one-to-one correspondence
between perfect model skill scores of deterministic and probabilistic
forecasts

TEI is the product of convective precipitation and storm relative
helicity (SRH), and here both factors are sensitive to the phases of
ENSO and the AO, with SRH showing stronger correlations (Figs. S6 and S7).

Although the model results here suggest a potential role for the joint
phases of ENSO and the AO in modulating severe thunderstorm activity,
a number of questions remain. Two key questions are whether the ENSO
and AO teleconnections in TEI found here are present in other climate
models and in reanalysis and whether relations with TEI translate to
relations with severe thunderstorm reports. These questions have been
explored for the ENSO signal

Sampling variability is a challenge to analyzing climate signals in
severe thunderstorm reports and reanalysis data. The teleconnection
patterns found here could provide guidance when pooling observational
data in time and space so as to reduce noise without diluting the
signal. For instance, the modest signals in April and May would
suggest that pooling data across the March–May season would be
suboptimal. In the same vein, analysis of observational data for
evidence of an AO signal may be more effective using daily data
because the persistence of the AO as measured by its autocorrelation
function tends to be less than 30 d

The cumulative distribution function (CDF) of a Gumbel-distributed random variable

Forecasts are of the probability of TEI exceeding its median value
conditional on the phases of ENSO and the AO. Suppose that during a
particular phase of ENSO and the AO, the Gumbel parameters of the TEI
distribution are

The power series approximation of

ONI data are provided by NOAA/CPC at

The supplement related to this article is available online at:

MKT designed the study, performed the analysis, made the figures, and wrote the manuscript draft. CL and MLL'H discussed the research and worked on revising the manuscript.

The contact author has declared that none of the authors has any competing interests.

Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This research has been supported by the Willis Research Network (grant no. WILLIS CU15-2366).

This paper was edited by Johannes Dahl and reviewed by Todd Moore and one anonymous referee.