atmosphere

Article
Potential Predictability of Seasonal Global Precipitation
Associated with ENSO and MJO
Haibo Liu 1 , Xiaogu Zheng 2, *, Jing Yuan 3 and Carsten S. Frederiksen 4,5,6

1 Lamont-Doherty Earth Observatory, Columbia University, Palisades, NY 10964, USA;

[email protected]
2 International Global Change Institute, Hamilton 3210, New Zealand
3 International Research Institute for Climate and Society (IRI), Columbia University, Palisades, NY 10964, USA
4 The Bureau of Meteorology, Melbourne, VIC 3001, Australia
5 The School of Earth, Atmosphere and Environment, Monash University, Clayton, VIC 3800, Australia
6 CSIRO, Oceans and Atmosphere, Aspendale, Melbourne, VIC 3195, Australia
* Correspondence: [email protected] or [email protected]

Abstract: A covariance decomposition method is applied to a monthly global precipitation dataset

to decompose the interannual variability in the seasonal mean time series into an unpredictable
component related to “weather noise” and to a potentially predictable component related to slowly
varying boundary forcing and low-frequency internal dynamics. The “potential predictability” is
then defined as the fraction of the total interannual variance accounted for by the latter component.
In tropical oceans (30◦ E–0◦ W, 30◦ S–30◦ N), the consensus is that the El Nino-Southern Oscillation
(ENSO, with 4–8 year cycles) is a dominant driver of the potentially predictable component, while the
Madden-Julian Oscillation (MJO, with 30–90 days cycles) is a dominant driver of the unpredictable
component. In this study, the consensus is verified by using the Nino3-4 SST index and a popular
MJO index. It is confirmed that Nino3-4 SST does indeed explain a significant part of the potential
predictable component, but only limited variability of the unpredictable component is explained by
the MJO index. This raises the question of whether the MJO is dominant in the variability of the
unpredictable component of the precipitation, or the current MJO indexes do not represent MJO
variability well.

Citation: Liu, H.; Zheng, X.; Yuan, J.;

Keywords: predictability; global; seasonal precipitation; ENSO; MJO
Frederiksen, C.S. Potential
Predictability of Seasonal Global
Precipitation Associated with ENSO
and MJO. Atmosphere 2023, 14, 695.
https://doi.org/10.3390/ 1. Introduction
atmos14040695 The seasonal mean time series of meteorological variables are widely used for ana-
Academic Editor: Seontae Kim
lyzing interannual climate variability and predictability. Studies of potential predictability
are usually based upon a decomposition of the seasonal variability into a part called the
Received: 17 February 2023 “weather noise” variability, which is fundamentally unpredictable beyond a deterministic
Revised: 2 April 2023 time period, and another part assumed to be at least potentially predictable [1]. The poten-
Accepted: 5 April 2023 tial predictability is measured as the fraction of the total variability accounted for by the
Published: 7 April 2023 latter part. It helps us to identify regions where the potential for making skillful climate
forecasts is highest, as well as where climate noise dominates any signal. Such knowledge
is useful to investigate the roles of drivers of potentially predictable variability, such as
human-induced forcing, oceanic and slowly varying atmospheric states, soil moisture and
Copyright: © 2023 by the authors.
snow cover conditions, and the roles of stochastic meteorological processes generated
Licensee MDPI, Basel, Switzerland.
This article is an open access article
by typical weather events in modifying regional variability. They can also help provide
distributed under the terms and
guidance as to where and when the issuance of climate forecasts may be most reasonable.
conditions of the Creative Commons The measure of potential predictability is clearly sensitive to how the separation of
Attribution (CC BY) license (https:// the variance is performed. Some may think that temporal filtering techniques could be
creativecommons.org/licenses/by/ employed based on the assumption that weather noise operates mainly on timescales
4.0/). much shorter than that of the potentially predictable variability [2]. However, weather

Atmosphere 2023, 14, 695. https://doi.org/10.3390/atmos14040695 https://www.mdpi.com/journal/atmosphere

Atmosphere 2023, 14, 695 2 of 12

events include not only high-frequency day-to-day fluctuations, but also intra-seasonal
fluctuations (such as MJO) which cannot be completely smoothed out by a seasonal mean
filter. The residuals through the seasonal mean filter give rise to chaotic, unpredictable
fluctuations in the seasonal mean time series. Therefore, it is not possible to completely
isolate the potentially predictable variability through temporal filtering.
In this paper, the methodology proposed by Zheng and Frederiksen [3] is applied to
analyze the potential predictability of a monthly global precipitation data. It was found
that for rainfall in tropical oceans, the main driver of predictable variability is ENSO,
while the unpredictable variability is related to MJO, but not as dominant as ENSO to the
predictable variability. Regions with significant potential predictability are also identified.
The methodology applied in this paper and other methodologies are reviewed.
The current paper begins with a description of the data and the estimation method
of the potential predictability in Section 2. Section 3 is devoted to applications to monthly
global precipitation data. The methodology applied in this paper and other methodologies
are reviewed in Section 4. A summary is presented in Section 5.

2. Data and Methods

2.1. Data
The data used in this study include monthly precipitation data (on a 2.5◦ × 2.5◦ grid)
from the Global Precipitation Climatology Project (GPCP), version 2.3 [4], for the period
1979−2020, which is a widely used global (land and ocean) precipitation dataset derived
from a mix of satellite estimates over oceans and land, and rain gauge measurements from
land and atolls. The GPCP is often used to study variations in precipitation at global
and regional scales (https://climatedataguide.ucar.edu/climate-data/gpcp-monthly-global-
precipitation-climatology-project#:~:text=The%20GPCP%20monthly%20dataset%20is,with%
20some%20delay%20for%20processing) (accessed on 6 April 2023). The Nino3-4 index is
calculated from the HadISST1 SST as the average of SST anomalies over the region 5◦ N–5◦ S
and 170◦ –120◦ W. The MJO index used here is the outgoing longwave radiation-based MJO
index (OMI) (https://www.psl.noaa.gov/mjo/mjoindex/) (accessed on 6 April 2023). The
OMI calculation is based on an empirical orthogonal function (EOF) analysis. The two com-
ponents of OMI, MJO-PC1 and MJO-PC2, are the projections of 20–96 days filtered outgoing
longwave radiation onto the first and the second daily spatial EOF patterns of 30–96 days
eastward filtered OLR. The OMI can successfully capture the convective component of the
MJO [5].

2.2. Identifying the Seasonal Predictable and Unpredictable Components

The (co-)variance decomposition method of Zheng and Frederiksen [3] is proposed for
deriving the spatial patterns of interannual variability in the seasonal mean fields related
to the variability of predictable and unpredictable components, based on monthly mean
data. First, the annual cycle is removed from the data by subtracting the multi-year mean
for each month. This serves to reduce the climate bias in the covariance estimation. Then, a
conceptual statistical model for a climate variable Xym in month m (m = 1, 2, 3 in a specific
season, e.g., June, July, August) and in year y (y = 1, . . . , Y, where Y is the total number of
years) expressed as per the following regression form:

Xym = µy + ε ym (1)

where µy (interception) represents the seasonal “statistical population” mean in year y; and
ε ym is the residual monthly departure of Xym from µy . Then, the seasonal mean of a climate
variable in year y (Xyo ≡ 13 ∑3m=1 Xym ) can be conceptually expressed as

Xyo = µy + ε yo (2)

where ε yo ≡ 13 ∑3m=1 ε ym is the seasonal mean of the monthly (also daily) residuals for µy
that arises from intra-seasonal variability (ε ym ) and is not smoothed out by the seasonal
Atmosphere 2023, 14, 695 3 of 12

mean filter. Since the weather system is chaotic, ε ym is unlikely to be predictable beyond
one or two weeks, so ε yo is essentially unpredictable beyond the deterministic prediction
period and is referred to as the unpredictable component of Xyo . On the other hand, the
interception µy is a constant through the season in y year and is more likely dominated
by the external forcing and slowly varying internal dynamics. Therefore, µy is potentially
predictable beyond the deterministic prediction period and is referred to as the potentially
predictable component of Xyo [1]. It is very difficult to numerically separate µy from the
seasonal mean Xyo using filter techniques, because Xyo is already the best estimation of µy ,
but subject
n to error ε yo . o
Let Xym0 , y = 1, . . . , Y; m = 1, 2, 3 be another climate variable in the form as Equation

(1), then, using the statistical technique based on the (co-)variance decomposition [4], the
interannual covariance of the unpredictable components can be calculated with monthly
data by using Equation (16) in Zheng and Frederiksen [3] as follows:
 
V ε yo , ε0yo = σ̂2 (3 + 4 ϕ̂)/9 (3)

where
a
σ̂2 = (4)
2(1 − ϕ̂)

ϕ̂ = min {0.1, max[( a + 2b)/(2a + 2b), 0]} (5)

1 1 Y  h 0 i 1 Y  h 0 i
a= { ∑ Xy1 − Xy2 Xy1 0
− Xy2 + ∑ Xy2 − Xy3 Xy2 0
− Xy3 } (6)
2 Y y =1 Y y =1

and
1 1 Y  h 0 i 1 Y  h 0 i
b= { ∑ Xy1 − Xy2 Xy2 0
− Xy3 + ∑ Xy2 − Xy3 Xy1 0
− Xy2 } (7)
2 Y y =1 Y y =1
 
The interannual covariance of the predictable components V µy , µ0y can be calculated
as      
V µy , µ0y = V Xyo , Xyo 0
− V ε yo , ε0yo (8)
 
0
where V Xyo , Xyo represents the total interannual covariance and can be estimated directly
0 , i.e.,
from seasonal mean variables Xyo and Xyo

  1 Y
0
V Xyo , Xyo = ∑ Xyo Xyo
0
(9)
Y y =1

The R-code for calculating the unpredictable covariance of Equation (3) is included in
Appendix A. The χ2 test and student’s t-test, used to judge the significance of the interan-
nual covariances of the predictable and unpredictable components, i.e., Equations (3) and (8)
respectively, are documented in the Appendix of Ying et al. ([6]; their Equations (10)–(14)).
The potential predictability is defined as the ratio of the predictable variance against
the total variance, i.e.,


V µy , µy /V Xyo , Xyo (10)
The significant test procedure for the potential predictability is documented in Equation (8)
of Zheng et al. [7].
Define the standardised predictable and unpredictable covariances for the Xyo 0 with

Xyo as
  r  
V µ0y , µy / V µ0y , µ0y (11)
Atmosphere 2023, 14, 695 4 of 12

and
  r  
V ε0yo , ε yo / V ε0yo, ε0yo (12)

respectively. It can be proved that the square of Equation (11) (Equation (12)) is the
variance of the predictable (unpredictable) component of Xyo explained by the predictable
(unpredictable) component of Xyo 0 (see Appendix B for proof). Therefore, the fraction of the

predictable (unpredictable) component of Xyo explained by the predictable (unpredictable)

component of Xyo 0 is the square of predictable (unpredictable) correlation.

     2
V2 µy , µ0y V µy , µ0y
 , cor2 (µy , µ0y )
 

  = r (13)
0 0
 
V µy , µy V µy , µy
 
V µy , µy V µ0y , µ0y

and  2
   
V2 ε yo , ε0yo V ε yo , ε0yo
2 0
 

  = r  =
 ˆ cor (ε yo , ε yo ) (14)
0 0
 
V ε yo , ε yo V ε yo , ε yo

V ε yo , ε yo V ε0yo , ε0yo

00 0 , then the
Moreover, if another climate variable Xyo is statistically independent of Xyo
fraction of the predictable (unpredictable) component of Xyo explained by the predictable
(unpredictable) components of Xyo0 and X 00 is the sum of those for X 0 and X 00 , respectively.
yo yo yo

3. Results
In principle, predictable variability is driven by the slowly varying external forcing
and slowly varying internal dynamics which are stable within a season, while unpredictable
variability is driven by the internal dynamics dominated by intra-seasonal variability with
cycles larger than 10 days. In this section, the impacts of ENSO and MJO on the predictable
and unpredictable variabilities of global precipitation, especially on tropical oceans, are
studied.

3.1. El Nino-Southern Oscillation

In this study, the Nino3-4 index is used to represent the ENSO. The potential pre-
dictability of the index is more than 0.9 for all seasons indicating that the predictable
variability is dominant. Therefore, any global precipitation related to ENSO is also likely to
be highly predictable.
Figure 1 shows that the variability in the predictable component of precipitation is
largest in the tropics in all seasons, especially in DJF. There is a clear seasonal cycle, with the
largest predictable variability in DJF and the smallest variability in JJA. This coincides with
the seasonal cycle of ENSO, which is also the strongest in DJF and the weakest in JJA. In the
peak season DJF, the center of the largest predictable variability is in the Nino3-4 region
(5◦ N–5◦ S, 170◦ W–120◦ W). This may be due to the fact that in the La Nina phase, the
cold equatorial water upwells in the Nino3-4 region, associated with a strong trade wind
passing through the Nino3-4 region, resulting in it being the driest; while in the El Nino
phase, the sea surface temperature in the Nino3-4 region is the warmest, resulting in a
large amount of convective cloud that leads to significant positive rainfall anomalies in
the Nino3-4 region. All these facts indicate that ENSO is a likely dominant driver of the
predictable variability of tropical precipitation. To further verify this point, the standardized
predictable covariances for the Nino3-4 index with global precipitation, in all four seasons,
are shown in Figure 2. Regardless of their signs, the spatial patterns and strength of their
absolute values are very similar to those of the predictable component variability (Figure 1)
in tropical oceans (30◦ E–0◦ W, 30◦ S–30◦ N). This is confirmed by the fact that the pattern
correlations between the predictable component standard deviations (Figure 1) and the
absolute values of the standardized predictable covariances for GPCP precipitation with
 standardized predictable covariances for the Nino3-4 index with global precipitation, in
all four seasons, are shown in Figure 2. Regardless of their signs, the spatial patterns and
strength of their absolute values are very similar to those of the predictable component
variability (Figure 1) in tropical oceans (30° E–0° W, 30° S–30° N). This is confirmed by the
Atmosphere 2023, 14, 695 fact that the pattern correlations between the predictable component standard deviations 5 of 12
(Figure 1) and the absolute values of the standardized predictable covariances for GPCP
precipitation with Nino3-4 SST (Figure 2) in the tropical oceans are 0.79, 0.82, 0.89 and 0.92
in MAM,
Nino3-4 JJA,(Figure
SST SON and DJF,
2) in therespectively.
tropical oceansTheare
corresponding means
0.79, 0.82, 0.89 andof0.92thein
fraction
MAM,ofJJA, the
predictable precipitation variances explained by the predictable component
SON and DJF, respectively. The corresponding means of the fraction of the predictable of Nino3-4
SST (the square
precipitation of Equation
variances (10))byover
explained the the tropical component
predictable oceans are 0.24, 0.41, 0.59,
of Nino3-4 SST 0.56, respec-
(the square
oftively. In SON
Equation (10))and
over DJF
thewhen ENSO
tropical is stronger
oceans are 0.24,and
0.41,unlikely to change
0.59, 0.56, phase,Inthe
respectively. SONcorrela-
and
tions
DJF andENSO
when explained predictable
is stronger variances
and unlikely to are more
change significant
phase, than thoseand
the correlations in explained
MAM and
JJA. Although
predictable the ENSO
variances variability
are more in MAM
significant is stronger
than those in MAM than
andthat
JJA.inAlthough
JJA, the correlation
the ENSO
and the explained predictable variance are less significant than those in JJA.
variability in MAM is stronger than that in JJA, the correlation and the explained predictable This coincides
with theare
variance factless
thatsignificant
ENSO is morethan likely
those to
inchange phase
JJA. This in MAM
coincides with(especially in April)
the fact that ENSO than
is
in JJA.
more likely to change phase in MAM (especially in April) than in JJA.

Atmosphere 2023, 14, x FOR PEER REVIEW 6 of 13

Figure1.1.Predictable
Figure Predictablecomponent
componentstandard
standarddeviation
deviationofofGPCP
GPCPprecipitation
precipitation(mm/day).
(mm/day).

Figure2.2.Standardized
Figure Standardizedpredictable
predictablecovariances
covariances for
for GPCP
GPCP precipitation
precipitation with
with Nino3-4
Nino3-4 SST.
SST.

3.2. Madden-Julian Oscillation

The Madden-Julian Oscillation (MJO) is a tropical weather system originating in the
Indian Ocean that propagates eastward around the global tropics with a cycle on the order
of 30–90 days [8]. The MJO indices MJO-PC1 and MJO-PC2 are the projections of 20–96
Atmosphere 2023, 14, 695 6 of 12

3.2. Madden-Julian Oscillation

The Madden-Julian Oscillation (MJO) is a tropical weather system originating in the
Indian Ocean that propagates eastward around the global tropics with a cycle on the order
of 30–90 days [8]. The MJO indices MJO-PC1 and MJO-PC2 are the projections of 20–96 days
filtered OLR [5], the seasonal potential predictability of the MJO-PC1 and MJO-PC2 are
virtual zero. Thus, the MJO is more likely to be a driver of the unpredictable variability of
tropical precipitation. Therefore, any global precipitation related to the two MJO indices is
also likely to be highly unpredictable on a seasonal time scale.
The MJO consists of two parts; one that has strong convective rainfall (wet) and the
other with suppressed rainfall (dry). The dry part of the MJO always precedes the wet part
(dry in the east, wet in the west). There are less clouds in the dry part, which induces strong
solar heating reaching the ocean surface underneath. The accumulation of the heating on
the ocean surface causes warmer SST, which in turn causes upward movement of the air.
This results in convection (wet part) moving eastward, and hence the east propagation of
the MJO. The MJO has wide ranging impacts on the patterns of tropical and extratropical
precipitation, atmospheric circulation, and surface temperature around the global tropics
and subtropics.
We compare the seasonal cycles of the MJO described by Zhang and Dong [8] with the
unpredictable variability of precipitation (Figure 3) and the standardized unpredictable
covariance for the MJO index with global precipitation (Figure 4) in tropical oceans. Zhang
and Dong [8] confirmed that the primary peak season is in boreal winter (December–March),
during which MJO signals are mainly confined to the Indian Ocean and western Pacific
Ocean, and reach their maxima in the South Pacific convergence zone. Similarly, Figure 2d
shows that the primary peak season of unpredictable variability is in DJF with the most sig-
nificant variability near the equatorial Indian Ocean (with standard deviation 1.5 mm/day)
and western Pacific Oceans, especially in the South Pacific convergence zone (with standard
deviation 2.0 mm/day). Figure 3a shows the unpredictable variability in MAM with a
similar pattern to DJF, but with less variability. This coincides with the fact that 7March
Atmosphere 2023, 14, x FOR PEER REVIEW of 13 also
belongs to the MJO peak season. This is also the case for the standardized unpredictable
covariances for GPCP precipitation with MJO index (Figure 4d) in the tropical ocean, but
withtropical
the much weaker strength
ocean, but with (with
muchstandard
weaker deviation 0.8–1.0
strength (with mm/day
standard in the Indian
deviation 0.8–1.0Ocean
and 0.6–1.0
mm/day mm/day
in the in the and
Indian Ocean South Pacific
0.6–1.0 convergence
mm/day zone).Pacific convergence zone).
in the South

Unpredictable
Figure3.3.Unpredictable
Figure component
component standard
standard deviation
deviation of GPCP
of GPCP precipitation
precipitation (mm/day).
(mm/day).
Atmosphere 2023, 14, 695 7 of 12

Figure 3. Unpredictable component standard deviation of GPCP precipitation (mm/day).

Figure
Figure 4. Standardized unpredictable
4. Standardized unpredictable covariances
covariances for
for GPCP
GPCP precipitation
precipitation with
with MJO
MJO index.
index.

Zhang and Dong [8] further confirmed that the secondary peak season of the MJO
Zhang and Dong [8] further confirmed that the secondary peak season of the MJO is
is boreal summer (June–September), during which the strongest MJO occurs north of the
boreal summer (June–September), during which the strongest MJO occurs north of the
equator from the Bay of Bengal to the South China Sea. Another separated region of
equator from the Bay of Bengal to the South China Sea. Another separated region of strong
strong MJO signals is located in the eastern Pacific warm pool off the Central American
MJO signals is located in the eastern Pacific warm pool off the Central American coast.
coast. Similarly, Figure 3b shows that the unpredictable variability in JJA also has a
Similarly, Figure 3b shows that the unpredictable variability in JJA also has a secondary
secondary peak north of the equator from the Bay of Bengal to the South China Sea (with
peak north of the equator from the Bay of Bengal to the South China Sea (with standard
standard deviation 1.5 mm/day) and a second center in the eastern Pacific warm pool
off the Central American coast (with standard deviation 1.0 mm/day). This is also the
case for the standardized unpredictable covariance for the global precipitations with the
MJO index (Figure 4b) for which JJA also has a secondary peak north of the equator
from the Bay of Bengal to the South China Sea, but with much weaker strength (with
standard deviation 0.8–1.0 mm/day). There is also a second center in the eastern Pacific
warm pool off the Central American coast, but with weaker strength (with standard
deviation 0.8 mm/day). Figure 3c shows the unpredictable variability in SON with a
similar pattern to that of JJA, but with less variability (Figure 3c), and this is also the
case for the standardized unpredictable covariance for the MJO index with tropical ocean
precipitations. This coincides with the fact that September also belongs to the second peak
season of MJO.
Zhang and Dong [8] suggested that for the broad tropical region, the seasonality
in the MJO is featured by a latitudinal migration across the equator between two peak
seasons. In particular, MJO signals in the eastern Pacific (Maloney and Kiehl [9]) appear to
be much stronger during boreal summer than winter. Such migration can be clearly seen
for its unpredictable variability. In JJA and SON (Figure 3b,c), the 1.0 mm/day contour
of the unpredictable standard deviation reaches north of Japan close to 50◦ N, while in
DJF it draws back to north of the Philippines about 15 N. Similarly, in DJF and MAM
(Figure 3a,d), the 1.0 mm/day contours reach north of New Zealand close to 30◦ S, while in
JJA it draws back to north of Australia about 10 S. However, these are not obvious in the
standardized unpredictable covariance for the MJO index with global precipitation. The
pattern correlations between the unpredictable component variability (Figure 3) and the
standardized unpredictable covariance of the precipitations with the MJO index (Figure 4)
in tropical oceans are 0.79, 0.79, 0.77 and 0.84 in MAM, JJA, SON and DJF, respectively. The
corresponding explained unpredictable variances by MJO-PC1 and MJO-PC2 are around
0.065 for all the seasons.
Atmosphere 2023, 14, 695 8 of 12

3.3. Potential Predictability

The spatial patterns of the potential predictability in the four seasons are shown in
Figure 5. In tropical oceans, they are more similar to the predictable variability (Figure 1)
than the unpredictable variability (Figure 2). In particular, the potential predictability
along the tropical Pacific is lowest in JJA (Figure 5b) associated with the lowest predictable
variability (Figure 1b). Potential predictability is highest in DJF (Figure 5d) associated with
the highest predictable variability (Figure 4d). Predictability along the tropical Atlantic is
Atmosphere 2023, 14, x FOR PEER REVIEW 9 of 13
the highest in MAM and lowest in SON (Figure 1a,c), which is consistent with the potential
predictability (Figure 5a,c).

Figure5.5.Seasonal
Figure Seasonalpotential
potentialpredictability
predictabilityofofGPCP
GPCPprecipitation.
precipitation.

Thepotential
The potential predictability
predictability of precipitations
precipitations in inthe
thesubtropics
subtropicsmay mayalsoalsobeberelated
related to
ENSO.
to ENSO. ForForexample,
example, thethe
JJAJJA
and DJF
and precipitation
DJF precipitation of New
of New Zealand
Zealand hashaspotential
potentialpredict-
pre-
ability around
dictability around0.3 (Figure
0.3 (Figure5b,d). Zheng
5b,d). and Frederiksen
Zheng and Frederiksen[10] further extended
[10] further the decom-
extended the
position methodology
decomposition methodologyfrom from
the variability of single
the variability time time
of single seriesseries
to thetocovariability of mul-
the covariability of
multivariate
tivariate time time series
series andand established
established a statistical
a statistical prediction
prediction scheme
scheme basedbased
on on
thethe predic-
prediction
tion of EOFs
of EOFs of the
of the covariance
covariance matrix
matrix ofofthe
thepredictable
predictablecomponents.
components. They They showed that thatthethe
predictability
predictabilityisisclosely
closelyrelated
relatedtotoNino3
Nino3SST,
SST,butbutalso
alsorelated
relatedtotothe
thetropical
tropicalIndian
IndianOcean
Ocean
SST,
SST,the
thelocal
localNZNZSSTSSTandandthethesouthern
southernannular
annularmode,
mode,withwiththethecross-validated
cross-validatedexplained
explained
variance
variance about 20%. For another example, the potential predictabilityofofthe
about 20%. For another example, the potential predictability theeastern
easternAsian
Asian
summer ◦ ◦
summermonsoon
monsoonrainfall
rainfall[the
[theseasonal
seasonalmeanmeanrainfall
rainfallininthe
theregion
region5–50
5–50°N, N,100–140
100–140°EEin in
June-July-August
June-July-Augustwhich whichincludes
includesthe theMeiyu
Meiyubeltbeltfrom
fromSouthern
SouthernJapan
Japanthrough
throughthe theYangzi
Yangzi
River
Rivercatchment]
catchment]isisaround
around0.3 0.3(Figure
(Figure5b).
5b).ByByusing
usingthethe methodology
methodology of of Zheng
Zheng and
and Fred-
Fred-
eriksen
eriksen [10], Yin et al. [11] demonstrated that the predictability is not only related tothe
[10], Yin et al. [11] demonstrated that the predictability is not only related to the
ENSO
ENSOdecaydecayandanddeveloping
developingphases,phases, butbutalso
alsototoseven
sevenother
otherpredictors
predictors fromfromthethePacific,
Pacific,
Atlantic
AtlanticandandIndian
Indian Oceans,
Oceans, making
making thethe
cross-validated
cross-validated explained variance
explained moremore
variance than 0.25,
than
which reaches about 80% of the potential predictability.
0.25, which reaches about 80% of the potential predictability.
Figure
Figure55also
alsoshows
showsthat thatthere
therearearesome
someareas
areaswith
withsignificant
significantpotential
potentialpredictability
predictability
in
in some seasons, for example, the western Mediterranean centered in northernEgypt
some seasons, for example, the western Mediterranean centered in northern Egyptin in
JJA (Figure 5b) and northern Mexico in DJF (Figure 5d). To
JJA (Figure 5b) and northern Mexico in DJF (Figure 5d). To our knowledge, no study our knowledge, no study
sim-
similar to Zheng
ilar to Zheng and
and Frederiksen
Frederiksen [10]has
[10] hasbeen
beencarried
carriedoutoutfor
forprecipitation
precipitation in in these
these regions.
regions.
However,
However, the potential predictability study in the current paper at least providesclues
the potential predictability study in the current paper at least provides cluesas as
to where the precipitation is more likely to be predicted and suggests that the statistical
prediction scheme proposed by Zheng and Frederiksen [10] is worth attempting.

4. Discussions on Methodology
Atmosphere 2023, 14, 695 9 of 12

to where the precipitation is more likely to be predicted and suggests that the statistical
prediction scheme proposed by Zheng and Frederiksen [10] is worth attempting.

4. Discussions on Methodology
In earlier studies, daily time series were used to estimate the unpredictable weather
noise component. A critical step is to separate the potentially predictable variability from
daily data. For continuous metrological variables, such as temperature and pressure, Mad-
den [1] applied a frequency domain approach where the potentially predictable variability
was removed by a Fourier transform on daily time series for all seasons. Shukla [12]
and Trenberth [13] applied an alternative time domain approach by assuming day-to-day
weather variability is red noise, and the potentially predictability component was partially
removed by a non-parametric approach in estimating the variability of the weather noise
component. The red noise constraint was further relaxed as colored noise, and the influence
of the potentially predictable component was completely removed by the application of the
first order difference operator (Zheng [14]) or modeled as a variance parameter (Jones [15])
or multi-year mean parameters (Delsole and Feng [16]). However, daily rainfall data are
intermittent and therefore the methodologies for daily continuous metrological data are
no longer suitable. Madden et al. [17] applied the chain-dependent model (Katz [18])
to estimate weather noise variability, but for the daily precipitation observations, the fit-
ted parameters are dependent on the predictable components (e.g., Katz and Zheng [19];
Zheng et al. [20]) and therefore the model is over simplified. Feng and Houser [21] mod-
eled the daily precipitation as Gaussian auto-regressive processes described by Delsole
and Feng [16], but the daily precipitation may not be Gaussian, especially if there are
considerable dry days in small areas.
Zheng et al. [7] developed a method for estimating the weather noise component only,
using monthly mean time series based on the likelihood estimation assuming the monthly
means of the weather noise components are Gaussian. They demonstrated, by simulation,
that the estimations of the weather noise component are comparable to those using daily
continuous meteorological data. Zheng and Frederiksen [3] extended the methodology of
Zheng et al. [7] for estimating the covariability of the potentially predictable components
based on the moment estimation with a relaxed Gaussian assumption. Since the Gaussian
assumption on the monthly mean of the unpredictable component is no longer required,
the methodology can, in principle, be applied to precipitation data.
In this paper, the methodology of Zheng and Frederiksen [3] was applied to analyze
the potential predictability of the GPCP precipitation. There are a number of advantages
of the methodology applied in this study over all the previous methodologies using daily
data. Monthly data are more available than daily data. This methodology can be applied
to estimate the covariability of the predictable (unpredictable) components between two
climate variables, which allows for the influence of one climate variable on another to be
estimated, as shown in this study. Moreover, it can be used to estimate the covariance
matrix of predictable (unpredictable) components of a climate field to which the singular
value decomposition analysis can be applied on the estimated component covariance
matrices. Such an approach was successfully applied to the seasonal prediction of New
Zealand rainfall (Zheng and Frederiksen [10]) and east Asia summer monsoon rainfall
(Ying et al. [11]). The successful predictions indicate that the estimation method based on
monthly rainfall data is reliable. Other existing methods can only be applied to estimate
the potential predictability of a single climate variable, but not the covariability between
two climate variables.
By using monthly data, fewer statistical assumptions are required compared with
methodologies using daily data. The only assumption is that the variance of ε ym is invariant
for month m, while the estimations using daily data depend on a variety of stationary
assumptions and parameterizations which can lead to quite different results. For example,
comparing the estimated potential predictabilities estimated in this paper (Figure 5) and
those using the methodology of Delsole and Feng [16] for the GPCP data (see Figure 1 of
Atmosphere 2023, 14, 695 10 of 12

Feng and Houser [21]), although there is much higher predictability for the South Pole
in their study, a higher predictability at the North Pole during JJA in our study is not
shown in their study. Furthermore, the potential predictability of the eastern Asia summer
monsoon rainfall is 30% using our method, while it is nearly zero in their study. Using our
method, Ying et al. [11] developed a seasonal prediction scheme with a 0.25 cross-validated
explained predictability for the eastern Asia summer monsoon rainfall.

5. Discussion and Conclusions

The covariance decomposition method proposed by Zheng and Frederiksen [3] was
applied to the monthly GPCP precipitation data. The variances of the precipitation seasonal
means for the four seasons were decomposed into potentially predictable components
driven by slowly varying boundary forcing and low-frequency internal dynamics, and
the unpredictable component related to day-to-day weather variability. It was suggested
that ENSO (represented by Nino3-4 SST) contributes more than half of the predictable
variability in SON and DJF when strong and unlikely to change phase. Although ENSO in
JJA is weaker than ENSO in MAM, it contributes about 40% of the predictable variability
in JJA compared with about 25% in MAM. This could be due to the fact that ENSO in
MAM is more likely to change phase, especially in April. The seasonal distribution of
the variability in the unpredictable component of GPCP precipitation is similar to that of
the MJO described by Zhang and Dong [8], though less than 7% of the variability in the
unpredictable component of GPCP precipitations was explained by the MJO index. This
raises the question as to whether the current MJO indexes represent MJO variability well,
or the MJO may not be so dominant in the variability of the unpredictable component
of seasonal precipitation compared to ENSO and, therefore, this subject is worth further
investigation.

Author Contributions: Conceptualization, H.L. and X.Z.; methodology, X.Z. and C.S.F.; formal
analysis, H.L., X.Z. and J.Y.; data curation, J.Y.; writing—original draft preparation, X.Z and H.L.;
writing—review and editing, C.S.F. All authors have read and agreed to the published version of the
manuscript.
Funding: This research received no external funding.
Data Availability Statement: The data used in this study are available publicly. The GPCP pre-
cipitation is available at: https://psl.noaa.gov/data/gridded/data.gpcp.html (accessed on 6 April
2023). The MJO index is available at: https://www.psl.noaa.gov/mjo/mjoindex/ (accessed on 6
April 2023).
Acknowledgments: We sincerely thank the reviewers for their comments to improve the paper.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
R-code for calculating the unpredictable covariance of Equation (4)
Cov.wmr=function(x1, x2) #x1,x2: Y by 3 matrix of the two monthly data sets
{
m <- dim(x1) [1]
X1 <- x1[, 1:2] − x1[, 2:3]
X2 <- x2[, 1:2] − x2[, 2:3]
a <- (sum(X1[, 1] * X2[, 1]) + sum(X1[, 2] * X2[, 2]))/2/m
b <- (sum(X1[, 1] * X2[, 2]) + sum(X1[, 2] * X2[, 1]))/2/m
alpha <- min(0.1, max((a + 2 * b)/(a + b)/2, 0))
sigma <- a/(1-alpha)/2
(sigma * (3 + 4 * alpha))/9
}
Atmosphere 2023, 14, 695 11 of 12

Appendix B
Proof of the variance of µy (ε yo ) explained by µ0y (ε0yo ).
We only prove for predictable component; proof for unpredictable component is
similar.
Suppose the numerical values of µy and µ0y exist, and µy can be regressed by µ0y as

µy = bµ0y + ξ y

where b is the regression coefficient and ξ y is the regression error. Since ξ y and µ0y are
statistically independent, then
   
V µy , µ0y = bV µ0y , µ0y

so the regression coefficient b can be estimated as

   
b = V µy , µ0y /V µ0y , µ0y

therefore, the variance of µy explained by µ0y is

       
V bµ0y , bµ0y = b2 V µ0y , µ0y = V 2 µy , µ0y /V µ0y , µ0y

Although the numerical values of µy and µ0y cannot be obtained, V (µy , µ0y ) can be
estimated using Equation (3).

References
1. Madden, R.A. Estimates of the natural variability of time averaged sea level pressure. Mon. Wea. Rev. 1976, 104, 942–952.
[CrossRef]
2. Basher, R.E.; Thompson, C.S. Relationship of air temperature in New Zealand to regional anomalies in sea-surface temperature
and atmospheric circulation. Int. J. Climatol. 1996, 16, 405–425. [CrossRef]
3. Zheng, X.; Frederiksen, C.S. Variability of seasonal-mean fields arising from intraseasonal variability: Part I, Methodology. Clim.
Dyn. 2004, 23, 177–191. [CrossRef]
4. Adler, R.F.; Huffman, G.J.; Chang, A.; Ferraro, R.; Xie, P.P.; Janowiak, J.; Rudolf, B.; Schneider, U.; Curtis, S.; Bolvin, D.; et al. The
version-2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor. 2003, 4,
1147–1167. [CrossRef]
5. Kiladis, G.N.; Dias, J.; Straub, K.H.; Wheeler, M.C.; Tulich, S.N.; Kikuchi, K.; Weickmann, K.M.; Ventrice, M.J. A comparison of
OLR and circulation based indices for tracking the MJO. Mon. Wea. Rev. 2014, 142, 1697–1715. [CrossRef]
6. Ying, K.; Frederiksen, C.S.; Zhao, T.; Zheng, X.; Xiong, Z.; Yi, X.; Li, C. Predictable and unpredictable modes of seasonal mean
precipitation over Northeast China. Clim. Dyn. 2018, 50, 3081–3095. [CrossRef]
7. Zheng, X.; Nakamura, H.; Renwick, J.A. Potential predictability of seasonal means based on monthly time series of meteorological
variables. J. Clim. 2000, 13, 2591–2604. [CrossRef]
8. Zhang, C.; Dong, M. Seasonality in the Madden–Julian Oscillation. J. Clim. 2004, 17, 1369–1380. [CrossRef]
9. Maloney, E.D.; Kiehl, J.T. MJO-related SST variations over the tropical eastern Pacific during Northern Hemisphere summer. J.
Clim. 2002, 15, 675–689. [CrossRef]
10. Zheng, X.; Frederiksen, C.S. A study of predictable patterns for seasonal forecasting of New Zealand rainfalls. J. Clim. 2006, 19,
3320–3333. [CrossRef]
11. Ying, K.D.; Jiang, D.; Zheng, X.; Frederiksen, C.S.; Peng, J.; Zhao, T.; Zhong, L. Seasonal predictable source of the East Asian
summer monsoon rainfall in addition to the ENSO-AO. Clim. Dyn. 2022, 60, 2459–2480. [CrossRef]
12. Shukla, J. Comments on “Natural variability and predictability”. Mon. Wea. Rev. 1983, 111, 581–585. [CrossRef]
13. Trenberth, K.E. Some effects of finite sample size and persistence on meteorological statistics. Part II: Potential predictability. Mon.
Wea. Rev. 1984, 112, 2369–2379. [CrossRef]
14. Zheng, X. Unbiased estimation of autocorrelations of daily meteorological variables. J. Clim. 1996, 9, 2197–2203. [CrossRef]
15. Jones, R.H. Estimating the variance of time averages. J. Appl. Meteor. 1975, 14, 159–163. [CrossRef]
16. Delsole, T.; Feng, X. The “Shukla–Gutzler” Method for estimating potential seasonal predictability. Mon. Wea. Rev. 2012, 141,
822–831. [CrossRef]
17. Madden, R.A.; Shea, D.J.; Katz, R.W.; Kidson, J.W. The potential long-range predictability of precipitation over New Zealand. Int.
J. Climatol. 1999, 19, 4. [CrossRef]
Atmosphere 2023, 14, 695 12 of 12

18. Katz, R.W. Precipitation as a chain-dependent process. J. Appl. Meteor. 1977, 16, 671–676. [CrossRef]
19. Katz, R.W.; Zheng, X. Mixture model for overdispersion of precipitation. J. Clim. 1999, 12, 2528–2537. [CrossRef]
20. Zheng, X.; Renwick, J.A.; Clark, A. Simulation of multisite precipitation using extended chain-dependent process. Water Resour.
Res. 2010, 46, W01504. [CrossRef]
21. Feng, X.; Houser, P. An examination of potential seasonal predictability in recent reanalysis. Atmo. Sci. Lett. 2014, 15, 266–274.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.