GEOPHYSICAL RESEARCH LETTERS, VOL. 23, NO. 1, PAGES 57-60, JANUARY 1, 1996
The 1990-1995
El Nifio-Southern
Oscillation
event:
Longest on record
Kevin E. TrenberthandTimothyJ. Hoar
NationalCenterfor Atmospheric
Research,
Boulder,Colorado
Abstract. The tendencyfor morefrequentE1 Nifio events
and fewer La Nifia eventssincethe late 1970's has been
linked to decad•l changesin climatethroughoutthe Pacific
basin. Aspectsof the most recentwarmingin the tropical
Pacific from 1990 to 1995, which are connectedto but not
Americato the InternationalDateline. It is the basin-scale
phenomenon,however,that is linked to globaJatmospheric
circulationand associatedweatheranomalies.The primary
responsein the atmospherecoupledto EN is the SO and,
together,the tropical Pacific warm eventsare often referred
synonymouswith E1 Nifio, are unprecedentedin the climate to as ENSO events.
record of the past 113 years. There is a distinction between
E1 Nifio (EN), the SouthernOscillation(SO)in the atmo- The
Southern
Oscillation
sphere, and ENSO, where the two are strongly linked, that
emergesclearly on decadal time scaJes.In the traditional E1
The SO has a time scaleof 2-7 yearsand consistsof a globalscale, predominantly standing wave with centersof action in
Nifio region,seasurfacetemperatureanomalies
(SSTAs)have
waxed and waned, while SSTAs in the centra• equatorial Pa- surfacepressureover Indonesia and the tropical South Pacific
cific, which are better linked to the SO, remained positive (seeFig. I of Trenberth and Shea1987). The nature of the
from 1990 to June 1995. We carry out severalstatisticaJtests oscillation can be seenfrom the inversevariations in pressures
at Darwin (12.4øS130.9øE)in northernAustraliaand Tahiti
(17.5øS149.6øW) in the South Pacificwhoseannualmean
part of a natural decadal-timesca•evariation. One test fits an
pressures
are correlatedat -0.79 (Trenberth1984). Thesetwo
autoregressive-moving
average(ARMA) modelto a measure
stationscan be optimally combinedinto a SO index, $0I -of the SO given by the first hundred years of the pressuresat
TN- DN whereT andD referto the departurefromthe long-
to assess the likelihood
that the recent behavior
of the SO is
Darwin, Australia, beginningin 1882. Both the recenttrend
term monthly mean sea level pressuresat Tahiti and Darwin,
for more ENSO events since 1976 and the prolonged1990respectively,and the subscriptN refers to an appropriate
1995 ENSO event are unexpectedgiven the previous record, normaJization.
with a probability of occurrenceabout once in 2,000 years.
Becausethere are many small scale and high frequency
This opensup the possibilitythat the ENSO changesmay be
phenomena.in the atmosphere that can influence the prespartly causedby the observedincreasesin greenhousegases.
suresat Tahiti and Darwin, such an index contains noisethat
doesnot reflect the global-scaleSO. A primary sourceof high
frequency variability arises from the 30-60 day intraseasonaJ
Introduction
The recent warm event related to E1 Nifio in the tropical
Pacific from 1990 to June 1995, is the longest on record since
1882 by some measures. It occurs in the context of a tendency for more frequent E1 Nifio events and fewer La Nifia
oscillation(Madden and Julian 1994). A measureof noise
SON - T•v 4- D•v (Trenberth1984)indicatestimeswhen
Tahiti and Darwin are not varying in opposition as observed
in the SO (Fig. 1). Taking accountof the signal-to-noise
ra-
tio and the coherenceof the Darwin and Tahiti time series,
the SOI is optimized as an indicator of the SO by applying
changesin climate throughoutthe Pacific basin (Trenberth a low-pass 11 term filter that eliminates fluctuations of less
1990, Trenberth and Huttell 1994, Kumar et al. 1994, Gra- than 8 monthsperiod but mostly retains periodsexceeding24
events since the late
1970's that
has been linked
to decad•l
ham 1995). In assessing
how unusualthe recentevent and months(Trenberth1984). The ratio of the variances
of these
the past two decadesare, due considerationmust be given to
two serieschangesfrom 2.0 for monthly data to 6.4 with the
whichregionsof the tropical'Pacific
seasurfacetemperatures filtering, and the filtered SOI retains56% of the monthlyvari(SSTs)are mostimportant,their link to the atmospheric
cir- ance. Accordingly, an appropriate normalization is the annual
culationcomponent
knownasthe SouthernOscillation(SO), meanstandarddeviationof eachtime series(Trenberth1984).
and the integrity and signaJ-versus-noise
ratio of any index or
measureof the SO. Given that the changesare unlikely in the
context of the natural variability as revealed by the previous
record, this may be evidenceof climate changerather than
simply an unusually long fluctuation.
The Tahiti recordis lessreliable and containsmissingdata
prior to 1935. Tahiti and Darwin monthly(annual)pressure
anomaliesare correlated-0.35 (-0.77) for 1935 to 1995 but
-0.19 (-0.36) for 1882to 1934yet the SO wasstronglyin evidencefrom otherstationsduringthe earlierperiod(Trenberth
E1Nifio (EN) refersto the occasional
%nomaJous"
warming and Shea1987). Th• only long-termhomogeneous
recordof
of the eastern tropical Pacific Ocean but is commonly linked the SO is givenby tie appropriatelyfilteredpressures
at Darto a basin-scalewarming extending from the coast of South win (Fig. 2). This recordis usedto judgethe degreeto which
the behavior in recent years is anomalous.
Copyright1996by theAmericanGeophysical
Union.
E1 Nifio-Southern
Papernumber95GL03602
Oscillation
To explore the coupling between the atmosphereand the
ocean in ENSO events, we make use of the "Nifio" regions
0094-8534/96/95 GL-03602503.00
57
$8
TRENBERTH
AND HOAR:
THE
1990-1995
ENSO EVENT:
LONGEST
ON RECORD
0.5
.0.5
-1.5
1.5
0.5
-0.5
-1.5
Raw noise
1.5
t
0.5
-0.5
:•
-3
1930
,,
i,
....
i,,,,1
1940
....
Smoothed
noise
i,,,
1950
1,,,,,
I,,,,,
....
1960
i,,,
1970
-1.5
....
,,
I,''''''''
1980
I ''1''-
1990
1055
Year
Figure 1. Time seriesof the SOI (normalizedTahiti minus
Darwin sea level pressureanomalies)signaland a measure
of noisein the SOI given by SON (normalizedTahiti plus
Darwin) with baseperiod 1950-1979.The top two showthe
SOI, monthly and filtered, the lower two showthe SON. The
filtered seriesreveal the interannual fluctuations;note the different scalingson the ordinate.
1960
1965
1070
1075
1980
1985
1990
1995
Year
Figure 3. Time seriesof Nifio areasSST anomaliesfrom
1950 through September 1995. The areas coveredare Nifio
[1+2]: 0ø-10øS 90øW-80øW, Nifio 3: 5øN-5øS, 150?W90øW, and Nifio 4: 5øN-5øS, 160øE-150øW.
tivelyregardedasindependent
thenthe 5%significance
levelis
0.30. Alsoapparent(Fig. 4) is the strikingboomerang-shaped
oppositeSST anomaliesextendinginto the extratropicsof
from the Climate DiagnosticsBulletin and SSTs from the each hemisphereand the coastalinfluencealong the AmerClimate AnalysisCenter. In the traditional E1 Nifio region, icaswith the samesignas the tropicalPacific. Therefore,the
Nifio [1+2] SST anomalies
(SSTAs)havewaxedandwaned, SSTs most involved in ENSO are in the central Pacific, not
while Nifio 4 SSTAs,like the SOI, haveremainedpersistently in the coastalwatersoff of Peru and Ecuador(Trenberthand
anomalous
in recentyears(Fig. 3).
Correlation coefficientsfor 1951 to April 1995 betweenthe
Shea1987,Deserand Wallace1987,Ropelewskiet al. 1992).
The importanceof the SO has variedin the past hundred
years: strong variations'from 1880 to the 1920sand after
smoothed
(rawmonthly)SOIandthemonthlySSTAs(Fig. 3)
are-0.61(-0.48),-0.81(-0.65)and-0.84(-0.68)forNifio[1+2], about 1950 and, except for strong event during 1939-1942,
weakervariationsfrom the mid-1920sto 1950(Trenberthand
Shea1987) (Fig. 2). Note the lackof decadalor longer-term
equatorialPacific. Nifio 4 and Nifio [1+2] SSTAsare corre- variability until recentyears. The trend for a negativeSOI aflated 0.49 over the same period. Commonly, SST increases ter 1976is apparent(Figs. I and 2) andsetsthe stagefor the
occurred almost simultaneously from South America to the very persistentwarmingoverthe past five yearsin the vicinity
Dateline} so the distinctionbetweenwhich regionmost influ- of the equatorand the Dateline (Fig. 3). It is importantto
encesthe atmosphere has not been easy to make. In recent distinguishthis phenomenonfrom the coastalE1 Nifio.
years,this has not beenthe case(Fig. 3). The correlations Usingthe smoothedSOI time seriesin Fig. 2, the previous
Nifio 3 and Nifio 4, respectively,demonstratinghighly statistically significantcorrelationsthat are strongestin the central
between the smoothed SOI and monthly SST anomaliesfrom
longest
continuous
ENSOevent,fromearly-1911
to mid-1915,
1950to 1994(45 years)(rig. 4) are strongest,
exceeding
-0.8, lasted 4.1 years. The longestinterval of oppositesign lasted
in the region120øWto 180,5øNto 10øS,whichstraddlespart 5.0 years,from May 1906to April 1911. The currentevent,by
of the Nifio 3 and Nifio 4 regionsand extendsfarther into the this measure,beganin October1989 and continuedthrough
Southern Hemisphere. If only one value per year is conserva- June 1995, 5 years 8 months. When the Nifio 4 SSTAs are
used as a measure,positive valuesbegan in September1989
andcontinuedthroughJuly 1995;valuesexceeded
0.3øCfrom
January1990throughJuly 1995exceptfor a shortbreakfrom
Februaryto April 1994(relativeto a 1950-79mean).
,o.
S•••••:•i•:...'-:•.'..•
[:!:
B•i:•:!:•-'.'•-!•
:•!-:•..
:i:
,
I
'
'"''"':•!...:'•i
i::i•ii*:'"•j•
::"'*':"':'•:'•:•.
•i::'.'!:i:!:i•'.-.: "
20N
Figure 2. Time seriesof the monthlynormalizedSouthern
Oscillation
Index(minusDarwinsealevelpressure
anomalies)
smoothedwith the l 1-term filter and with a low passsmoothing spline filter that removesperiodslessthan 10 years. A
1-2-1 filter is used on the end points and the last value is
September 1995.
40SI00E
120[140[160[ 180 160W
140W
120W
100W
60W60W'
Figure 4. Correlationsof the smoothedSOI with monthly
SST anomaliesfor January 1950 to December1994 (540
months).Magnitudesexceeding0.3 are shadedor stippled.
TRENBERTHAND HOAR:THE 1990-1995ENSOEVENT: LONGESTON RECORD
Statistical
59
tests
To test whether the recent SO behavior is statistically un-
likely,a numberof testsare performed.We identifythe seasonsas December-January-February,
DJF, and so on. The 22
I ! I If I I I I
seasonalmean anomalies at Darwin from DJF 1989 to MAM
1995are all positiverelativeto the 1882-1981mean. The next
longestpositiverunsobserved
are15seasons
fromMAM 1894
to SON 1897,12 fromDJF 1981-82to SON 1984,and 11 from
0
50
1O0
150
200
250
300
350
400
Season
SON 1939 to MAM 1942. The longest run of oppositesign Figure 6.
lasted 13 seasons from MAM 1973 to MAM 1976.
We first use a t test for the difference between the means of
Observed
(top) from 1882to 1981and simu..
lated(bottom)100yearsequences
of Darwinseasonal
pressure anomalies. Note the additional noisecomparedwith the
two periodsunderthe null hypothesis
that they arethe same, filtered values in Fig. 2.
modifiedto accountfor the strongpersistencein the SO time
series.In determiningthe standarderror of the meanfor the
SOI, the time neededto gain an extra degreeof freedomor, wherethe at is a Gaussianwhite noiseprocessat time t with
effectivelythe time betweenindependentobservations,is 6 zero mean and variance era
2 _ 0.435 mb2, andthe y is the
months(Trenberth1984). For December1989to May 1995 syntheticDarwinseriesof seasonal
anomalies,
t is in unitsof
(66 months)versusJanuary1882to November
1989(1295 seasons,and all units are in mb.
months),the resultis t - 2.82with an estimated225degrees The estimatedpower spectrumof the Darwin seasonal
of freedom,significantat 0.5% with a two-tailedtest.
anomalies
for 1882-1981(Fig. 5), whichrevealsa broadspecSimilarly,for March 1977-May 1995 (219 months),the t tral peakfrom3 to 7 yearsperiodin the SOI, is seento havea
statistic is 2.68; significantlydifferentfrom the mean for the verygoodfit with the theoretical
ARMA spectrum
(c.f. Chu
1142 months in the rest of the series at • 1%.
andKatz 1989),whichfeaturesa broadpeakat 4.2 yearspeThese tests provide a ballpark indication, but are con- riodicity. Of particularnote for the currentexercise,is the
founded somewhat by the persistencein the time series. A excellentagreement
in powerat very low frequencies.This
more comprehensive
test fits an autoregressive
(AR) mov- modelis usedto generateover 1,000,000yearsof synthetic
ing average(MA) or ARMA modelto the time seriesof sea- recordand the first 1,000yearsare discardedbecauseof possonalDarwin pressureanomaliesfor the first hundredyears siblespinupeffects. The remainderis usedto estimatethe
of record(1882to 1981). The modelis then usedto generate probabilityof eventssuchasthoserecentlyobserved.
A comsynthetic time seriesfor testing.
parisonof 100 yearsof modeloutputwith the observed
first
The ARMA model for Darwin can be compared with one 100 yearsof Darwin (Fig. 6) revealsthe similarcharacterof
based on a much shorter period for the SOI Tahiti minus the variations;notethreesequential
strongsimulated"ENSO
Darwin by Chu and Katz (1985) who found a best fit with events"3.5 yearsapart betweenseasons200 and 250, quite
an AR(3) processfor seasonal
values.Our bestfit modelwas similar to the observed1939-42event (near season240, top
foundusinga maximumlikelihoodfitting procedure(Jones curve).
1980)andAkaike's
Information
Criterion(AIC)(Akaike1974)
Because the results indicated we were dealing with rare
which invokesa performanceand penalty functionrelatedto
the number of parameters in the model; the most parsimo-
events,we testedthe sensitivityof the resultsto variantsin
the model. Minor changesarisingfrom differentinitial states
nious model is one with a minimum
in the search for the ARMA coefficientsmade insignificant
in the AIC.
The model
selectedis an ARMA (3,1) (of order3 for the AR process
and
1 for the MA), givenby
Yt -- 1.278yt_• - 0.318yt-2 - 0.136yt_a+ at - 0.712at_1
Period
101 ß/ /
- Seasons
,'-- 1•;00 ,5.?0 ;•.,3;• 2.,50 2.00
•
•
Bandwidth
differences.Somewhatlarger differences
wereproducedwhen
the model was based on either the first 95 or 105 years of
record instead of the first 100. In both cases,the spectrumof
the ARMA best-fit model had slightlymorepowerat periods
longerthan 20 years. The 105-yearmodelis basedon data
through1986 whichincludesthe 1982-83ENSO event, the
largeston record. Becauseit overlapsconsiderably
with the
1977-1995periodthat we wish to test, it is considered
less
reliable.
\l \•, /- /'•
10
• /
\/
\
..-.•
---/
/'\
\/
I
'
The longestrun of simulatedvaluesof one sign was 31
seasonsand 113 .runs of either sign equalledor exceeded22
seasons,
indicatingan incidenceof 1 "observed"
eventin about
8,850 years. From this standpoint,the model suggeststhat
the continuouspositiverun for the most recent22 seasonsis
extremelyrare. In the alternatemodels,the numbersof runs
of 22 seasonsor more were 144 and 154, so that the most
Frequency- Cycles/Season
Figure 5. Estimatedpowerspectrafor Darwinseasonal
conservative estimate is for one such run about every 6,500
years.
anomaliesfrom 1882-1981(heavy) and the corresponding The mean Darwin anomaly relative to 1882-1981 for the
spectraof the ARMA(3,1)bestfit model(thin) . A loga- the last 22 seasonsis 0.94 mb. From the model, it was deterrithmic vertical scaleis usedso that the 95% confidenceband
mined that the median 1 in 1000 year threshold for 22-season
(dashed)
is constant
in width,whiletheabscissa
is linearin eventsof either sign was 0.88 mb, and the 1 in 10,000 year
frequency.
The bandwidth
of the Parzenwindow(with 60 threshold .was0.99 mb, indicating a frequencyof occurrence
lags)is plotted.
of the observedvalue of 1 in 3,000 years; or about 1 in 545
60
TRENBERTH AND HOAR: THE 1990-1995ENSO EVENT.: LONGEST ON RECORD
22-season
intervals(0.18%). Given100trials of 10,000years, in greenhousegasesin the atmosphere?Or is this pattern
the samplingfor 1,000year eventsindicatesan interquartile natural decadal-timescale variation? We have shown that the
rangeof 0.86 to 0.91 mb, and the rangeis 0.80 to 0.99 mb. The latter is highly unlikely.
median return periods of the observedvalue in the alternate
modelswere 1,648 and 1,137 years.
Acknowledgments.
NCAR is sponsored
by the National
For the last 73 seasons
from MAM 1977to MAM 1995,the Science Foundation.
We thank Dennis Shea and Chris Guillemot
observedmeananomalyis 0.50mb. From73-season
sequences for help with the figures and datasets, and Richard Jones for help
in the model, valuesof 0.48 mb and 0.52 mb are reachedonce in fitting the ARMA models. This researchis partly sponsoredby
every 1,000 and 10,000 years, respectively,and the observed NOAA under grant NA56GP0247 and by the GeophysicalStatistics
valueis obtainedabout onceevery 2,000 yearsor about I in Project at NCAR under grant NSF DMS9312686.
110 suchsequences
(0.9%). The resultsfrom the alternative
References
modelsgivereturn periodsof 2,400 and 1,390years.
In all cases,the shortestreturn periodsoccurfor the model Akaike, H., A new look at the statistical model identificabasedupon 1882-1986,as expected.For the 73-season
case, tion. IEEE Trans. Auto. Control, 19, 716-723, 1974.
the longestreturn periodis for the modelbasedupon 1882- Chu, P-S., and R. W. Katz, Modelingand forecastingthe
1976,the onlymodelthat wasnot basedonpart ofthe period
Southern Oscillation: A time-domainapproach. Mon.
of interest.
Wea. Rev. 113 1876-1888, 1985.
Chu, P-S., and R. W. Katz, Spectralestimationfrom time
series models with relevance to the Southern Oscillation.
Discussion
J. Climate 2, 86-90, 1989.
The resultsfromthe t testsindicatethat the lowfrequency Deser C., and J. M. Wallace, E1 Nifio events and their revariability and the negativetrend in the SOI in recentdecades
lation to the SouthernOscillation.J. Geophys.Res.,92,
14189-14196, 1987.
are quiteunusual;the recentsequence
wouldbe expectedonly
aboutoncein 1,100years.The ARMA modelingrefinesthis Graham, N. E., Simulationof recent global temperature
estimate and indicates that the likelihood of occurrence of
trends. Science,267, 666-671, 1995.
the 1990-95ENSO, gi•:enthe first hundredyearsof record, Jones,R. H., Maximum likelihoodfitting of ARMA models
is aboutoncein 1,500to 3,000years.The changebeginning to time serieswith missingobservations.Technometrics,
22, 389-395, 1980.
about 1976 toward more frequentENSO eventsis also unlikely,abouta oncein 2,000yearexpectation.Clearlythe two Knutson,T. R., and S. Manabe, Time-meanresponseover
are not independentand the resultssuggesta non-stationary the tropicalPacificto increasedCO2 in a coupledoceaninfluencesuchas wouldarisefrom a changein climate.
atmospheremodel. J. Climate 8, 2181-2199, 1995.
Severalaspectsof the decadalchangebeginningaround Kumar, A., A. Leetmaa and M. Ji, Simulations of atmo1976 have been simulatedwith atmospheric
modelsusing
sphericvariability inducedby sea surfacetemperatures
and implicationsfor globalwarming. Science,266, 632specifiedSSTs (Kumar et al. 1994, Graham1995)includ634, 1994.
ing patternsof surfacetemperaturechangeobserved
in recent
years over land. These and other studiesconfirm that the
atmosphericchangesare tied to the changes
in SSTsand, fur-
Madden, R. A., and P. R. Julian, Observationsof the 40-50
daytropicaloscillation-A review. Mon. Wea. Rev. 122,
ther,that the changes
overthe NorthPacificandsurrounding
814-837, 1994.
areas are substantially controlled by the anomalousSST forc- Ropelewski,
C. F., M. S. Halpert, and X. Wang,Observed
ing from the tropicalPacific,and thuslinkedto the changes troposphericbiennialvariabilityandits relationshipto the
in ENSO. Accordingly,
the key questionis whythe SSTs,and
SouthernOscillation. J. Climate5, 594-614, 1992.
especiallythe tropicalPacificSSTs,havechangedin the way Trenberth, K. E., Signal versusnoisein the SouthernOscilthey have?
lation. Mon. Wea. Rev., 112, 326-332, 1984.
No currentcoupledglobalclimatemodelprovidesa fully Trenberth, K. E., Recent observed interdecadal climate
realisticsimulationof ENSO eventsalthoughsomemodelsrechangesin the Northern Hemisphere. Bull. Amer. Meproduce aspectsof the observedbehavior but with reduced
teor. Soc., 71,988-993, 1990.
amplitude.Thusthereis little basisfor expecting
that there Trenberth,K. E., and J. W. Hurrell,Decadalatmospherewill be an increasein the numberand/or lengthof ENSO
oceanvariationsin the Pacific. Clim. Dyn., 9, 303-319,
1994.
eventsfromcoupledatmosphere-ocean
models,although
modelsdo predictSSTwarmingwith increased
greenhouse
gases Trenberth, K. E., and D. J. Shea, On the evolution of the
(KnutsonandManabe1995),sothat the background
mean SouthernOscillation.Mon. Wea. Rev.,115, 3078-3096,
1987.
is changed.In morelimitedmodelsdesigned
specifically
for
simulatingENSO in the Pacific,the prolongedENSO hasnot
been well predicted by models which have otherwisedemonstratedskill in predictingENSO events(seerecentissuesof
NOAA's Experimental long-lead Forecast Bulletin and Climate DiagnosticsBulletin).
These results raise questions about the role of climate
change.Is this patternof changea manifestationof the global
warmingand related climate changeassociatedwith increases
K. E. Trenberth
and T. J. Hoar.
Climate
and Global
Dynamics Division, National Center for Atmospheric
Research, P.O. Box 3000, Boulder, CO 80307-3000.
(e-mail: [email protected])
(received
August,21, 1995;revised
October9, 1995;
acceptedNovember13, 1995.)