Kim, Namhun; Cha, Junhee; Jeon, Junwoo
A icle
A compa a i e e alua ion o machine lea ning app oaches
o con aine eigh a es p edic ion
Asian Jou nal o Shipping and Logis ics (AJSL)
P o ided in Coope a ion wi h:
Ko ean Associa ion o Shipping and Logis ics, Seoul
Sugges ed Ci a ion: Kim, Namhun; Cha, Junhee; Jeon, Junwoo (2025) : A compa a i e e alua ion o
machine lea ning app oaches o con aine eigh a es p edic ion, Asian Jou nal o Shipping and
Logis ics (AJSL), ISSN 2352-4871, Else ie , Ams e dam, Vol. 41, Iss. 2, pp. 99-109,
h ps://doi.o g/10.1016/j.ajsl.2025.05.001
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/329761
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
A compa a i e e alua ion o machine lea ning app oaches o con aine
eigh a es p edic ion
Namhun Kim
a
, Junhee Cha
a
, Junwoo Jeon
b,*
a
Depa men o Business Adminis a ion, Sungkyul Uni e si y, Republic o Ko ea
b
Depa men o Global Logis ics, Sungkyul Uni e si y, Republic o Ko ea
ARTICLE INFO
Keywo ds:
Con aine F eigh Ra es
Machine lea ning
Decision T ee
Random Fo es
LSTM
P ophe
ABSTRACT
This s udy e alua es he p edic i e pe o mance o ou models—Decision T ee, Random Fo es , P ophe , and
LSTM—in o ecas ing con aine eigh a es, a key me ic o s a egic decision-making in he shipping indus y.
To add ess da a he e ogenei y, Min-Max no maliza ion was applied, and he Johansen co-in eg a ion es
con i med long- e m ela ionships among he a iables, jus i ying he use o aw da a in ou analysis. Pe o -
mance was assessed using MSE, RMSE, NMSE, MAE, MAPE and SMAPE. While bo h Decision T ee and Random
Fo es models yielded lowe absolu e e o s compa ed o LSTM and P ophe , he Decision T ee model demon-
s a ed supe io ela i e accu acy, ou pe o ming Random Fo es by app oxima ely 91.8 % on he USWC ou e,
52.1 % on USEC, 43.5 % on MED, and 22.7 % on NEUR. These indings highligh he obus ness o he Decision
T ee model o con aine eigh a e o ecas ing unde ola ile ma ke condi ions.
1. In oduc ion
The shipping indus y is a co ne s one o global ade, esponsible o
anspo ing o e 80 % o he wo ld’s ca go. This i al unc ion un-
de pins global economic g ow h and supply chain s abili y (Wang e al.,
2024). In his con ex , he accu a e p edic ion o ocean eigh a es is
c i ical. F eigh a es no only e lec cu en ma ke condi ions bu also
se e as a mechanism o balancing supply and demand h ough s a-
egic a e adjus men s (Jeon e al., 2020; Sca si, 2007; Sch amm &
Munim, 2021; Wang e al., 2024).
The impo ance o o ecas ing ocean eigh a es is mul i ace ed.
Fi s , key s akeholde s—including ship owne s, ca go ca ie s, logis ics
companies, and end consume s— ely on hese p edic ions o in o med
decision-making ac oss a ious unc ions such as p oduc p icing, cos
accoun ing, inancial managemen , and asse alloca ion (Hi a a &
Ma suda, 2022; Jeon e al., 2021). The ola ili y o eigh a es is a
undamen al componen o ma i ime anspo cos s; he e o e, p ecise
o ecas ing o his ola ili y is c ucial o ma ke pa icipan s (Naima
e al., 2023). Second, he inhe en pe iodici y o ocean eigh a es
a ises om he leng hy in e al be ween ship o de s and deli e ies,
which ypically exceeds wo yea s. This empo al lag ein o ces cyclical
pa e ns in eigh a e dynamics (Sca si, 2007). Thi d, ship owne s
make ou ine decisions ega ding ship sales and cha e s based on
p e ailing eigh a e le els, u he emphasizing he need o obus
o ecas ing (Jeon e al., 2020). Fou h, o shippe s, accu a e a e p e-
dic ions a e essen ial as logis ics cos s a e di ec ly co ela ed wi h eigh
a es—cos s end o escala e du ing pe iods o high a es and decline
when a es a e low. Addi ionally, seasonal a ia ions, exempli ied by he
implemen a ion o Peak Season Su cha ges (PSS) in mon hs such as
Ma ch o Oc obe , u he complica e he o ecas ing landscape (Yin &
Shi, 2018). Finally, ex e nal shocks, including geopoli ical isks and
e en s such as he Suez Canal blockage du ing he COVID-19 pandemic
and he ecen Red Sea c isis, ha e demons a ed signi ican ad e se
impac s on con aine eigh a es. Inco po a ing hese dis up ions in o
o ecas ing models has been shown o enhance p edic i e accu acy
(Naima e al., 2023).
Mo eo e , he shipping indus y is no ably capi al-in ensi e, wi h
subs an ial inancial commi men s equi ed o ship o de ing and
ope a ion (Jeon & Yeo, 2017). O e capaci y, o en esul ing om
agg essi e shipbuilding, can p ecipi a e a decline in eigh a es,
he eby in ensi ying he need o accu a e ma ke assessmen s and isk
e alua ions. Fo ins ance, using sys em dynamics, Jeon e al. (2020)
iden i ied a cyclical pa e n o app oxima ely 32 mon hs in he China
Con aine F eigh Index (CCFI), highligh ing he in e play be ween
supply-demand imbalances. Addi ionally, s ingen en i onmen al eg-
ula ions imposed by he In e na ional Ma i ime O ganiza ion (IMO),
* Co esponding au ho .
E-mail add ess: [email p o ec ed] (J. Jeon).
Con en s lis s a ailable a ScienceDi ec
The Asian Jou nal o Shipping and Logis ics
jou nal homepage: www.else ie .com/loca e/ajsl
h ps://doi.o g/10.1016/j.ajsl.2025.05.001
Recei ed 21 Ma ch 2025; Accep ed 6 May 2025
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
A ailable online 17 May 2025
2092-5212/© 2025 The Au ho (s). Published by Else ie B.V. on behal o The Ko ean Associa ion o Shipping and Logis ics, Inc. This is an open access a icle unde
he CC BY license ( h p://c ea i ecommons.o g/licenses/by/4.0/ ).
such as he Ene gy E iciency Design Index (EEDI), Ene gy E iciency
Ship Index (EEXI), and Ca bon In ensi y Indica o (CII), a e compelling
shipping companies o in es in eco- iendly essels. These egula o y
equi emen s u he ampli y he necessi y o eliable eigh a e
o ecas s o mi iga e he isks associa ed wi h new ship o de s
(Bay ak a & Yuksel, 2023; Lee, 2024).
In ligh o hese conside a ions, his s udy aims o compa e he p e-
dic i e pe o mance o ou dis inc models—P ophe , Decision T ee,
Random Fo es , and LSTM—in o ecas ing con aine eigh a es. The
objec i e is o iden i y he model ha mos e ec i ely cap u es he
complex dynamics o ocean eigh a es, he eby suppo ing mo e
in o med decision-making by indus y s akeholde s.
The emainde o his pape is o ganized as ollows. Chap e 2 e-
iews ele an li e a u e on con aine eigh a e p edic ion and dis-
cusses p io applica ions o he P ophe , Decision T ee, Random Fo es ,
and LSTM models. Chap e 3 de ails he da a and me hodological
amewo k employed. Chap e 4 p esen s empi ical indings and p o-
ides a compa a i e e alua ion o he models. Chap e 5 discusses he
implica ions o he indings, and Chap e 6 concludes he s udy.
2. Li e a u e e iew
2.1. Fo ecas ing s udies on con aine eigh a es
Con aine eigh a es a e c i ical indica o s o a mul i ude o
s akeholde s wi hin he shipping indus y, in luencing decisions anging
om p icing and inancial managemen o asse alloca ion. T adi ion-
ally, econome ic models—such as ARIMA, VAR, VEC, and
ARMA—ha e been ex ensi ely employed o o ecas hese a es
(Koyuncu & Ta acıo˘
glu, 2021; Munim & Sch amm, 2017). Fo example,
Sch amm and Munim (2021) conduc ed compa a i e analyses be ween
ARIMA and VAR, while Munim and Sch amm (2021) ex ended his
compa ison o ARIMA and VEC. In ano he s udy, Chen e al. (2021)
p oposed an inno a i e hyb id app oach by in eg a ing empi ical mode
decomposi ion (EMD) wi h ARMA. Luo e al. (2009) u he ad anced
he ield by employing supply and demand dynamics o o ecas ing
con aine eigh a es.
No wi hs anding hese con ibu ions, he inhe en ly non-linea and
complex na u e o con aine eigh a e da a p esen s subs an ial
challenges o adi ional econome ic models (Hi a a & Ma suda,
2022). This limi a ion has ca alyzed a shi owa ds he adop ion o
machine lea ning and deep lea ning echniques, which a e mo e adep
a cap u ing non-linea pa e ns. Recen s udies ha e demons a ed ha
deep lea ning models, such as LSTM, can ou pe o m con en ional ap-
p oaches in ce ain con ex s; o ins ance, Hi a a and Ma suda (2022)
epo ed supe io p edic i e pe o mance o LSTM o e ARIMA o
deep-sea ou es, al hough esul s we e compa able o sho -sea ou es.
Mo eo e , Chen e al. (2024) showed ha a CNN-LSTM hyb id model
could achie e an R² exceeding 90 %, ou pe o ming an ARIMA-SVR
amewo k. Simila ly, machine lea ning models such as Random Fo -
es ha e been shown o deli e p edic ion accu acy abo e 80 % (Feng,
2022; Khan & Hussain, 2022). Complemen a y wo k by Jeon e al.
(2021) compa ed ARIMA, VEC, and Sys em Dynamics models, inding
ha he la e educed p edic ion e o by app oxima ely 30 % on
a e age, he eby o e ing dis inc ad an ages in cap u ing ma ke
dynamics.
2.2. S udies on speci ic o ecas ing models
In addi ion o b oad econome ic app oaches, a ious o ecas ing
models ha e been applied ac oss di e en domains. Thei ele ance o
con aine eigh a e p edic ion is elucida ed below.
2.2.1. Decision ee models
Decision ee me hodologies ha e p o en e ec i e in nume ous
o ecas ing applica ions. Liu e al. (2017a) applied decision ees o
coppe p ice p edic ion, achie ing obus pe o mance as measu ed by
MAPE and RMSE. Bala (2010) ex ended his applica ion o demand
o ecas ing o Indian e aile s, demons a ing ha decision ee-based
models ou pe o med o he me hods—such as ARIMA and SARIMA—-
o e bo h sho and long- e m ho izons. Hyb id models ha combine
decision ees wi h a i icial neu al ne wo ks (ANN) ha e u he
enhanced p edic ion accu acy (Chang, 2011; Tsai & Wang, 2009),
ein o cing he e sa ili y o decision ee amewo ks.
2.2.2. Random o es models
Random Fo es , an ensemble ex ension o decision ees, mi iga es
he isk o o e i ing and enhances p edic i e s abili y. Liu and Li
(2017b) u ilized Random Fo es o o ecas gold p ice luc ua ions,
iden i ying c i ical p edic o s such as he DJIA and S&P 500 indices.
Kuma and Thenmozhi (2006), demons a ed ha Random Fo es p o-
ided compe i i e accu acy ela i e o SVM, LDA, and logis ic eg ession
o p edic ing s ock ola ili y. Subsequen s udies by Ab aham e al.
(2022) and Vai agade e al. (2019) ha e consis en ly epo ed ha
Random Fo es ou pe o ms al e na i e me hods in a ious o ecas ing
con ex s. Hue as Ta o and Cen eno B i o (2018), u he alida ed hese
indings h ough applica ions in sola ene gy p oduc ion o ecas ing,
while Xue e al. (2021) compa ed mul i-objec i e Random Fo es a i-
an s, con i ming i s obus ness in handling complex p edic ion asks.
2.2.3. LSTM models
Long Sho -Te m Memo y (LSTM) ne wo ks, a subse o ecu en
neu al ne wo ks (RNN), a e pa icula ly well-sui ed o ime-se ies
o ecas ing due o hei abili y o cap u e long- e m dependencies.
Bhanda i e al. (2022) ound ha single-laye LSTM models p o ided
supe io p edic i e accu acy o S&P 500 closing p ices when e alua ed
agains RMSE, MAPE, and R² me ics. Con e sely, Abbasimeh e al.
(2020) epo ed ha mul i-laye LSTM models achie ed e en g ea e
pe o mance, ou pe o ming con en ional models such as ARIMA, ANN,
and SVM. Saghee and Ko b (2019) ex ended he LSTM a chi ec u e
(DLSTM) o pe oleum p oduc ion o ecas ing, demons a ing no able
imp o emen s in RMSE and MAPE, while Siami-Namini e al. (2018)
documen ed signi ican e o educ ions o 84–87 % in compa ison o
ARIMA models.
2.2.4. P ophe models
P ophe , de eloped by Taylo and Le ham (2018), is designed o
obus ime-se ies o ecas ing by accommoda ing ends, seasonali y,
and ex e nal e en s. Empi ical compa isons by Jha and Pande (2021)
e ealed ha P ophe achie ed lowe RMSE and MAPE alues compa ed
o ARIMA in he con ex o supe ma ke sales o ecas ing. Simila ly,
Yenido˘
gan e al. (2018) applied bo h models o Bi coin p ice p edic ion,
epo ing an R² o 0.94 o P ophe e sus 0.68 o ARIMA. Kaninde e al.
(2022) u he demons a ed he u ili y o P ophe in ola ile s ock
ma ke o ecas ing, pa icula ly due o i s abili y o in eg a e holiday
e ec s and o he exogenous ac o s. Recen ex e nal shocks, including
he Suez Canal blockage du ing COVID-19 and he Red Sea c isis, un-
de sco e he impo ance o inco po a ing exogenous a iables in o
o ecas ing models o enhance p edic i e accu acy (Naima e al., 2023;
Wang e al., 2024).
2.2.5. Con ibu ions
Despi e he ex ensi e body o wo k u ilizing econome ic models o
con aine eigh a e p edic ion, he e emains a signi ican gap in he
applica ion o ad anced machine lea ning and deep lea ning echniques
o his domain. In pa icula , he ela i e unde u iliza ion o models
such as P ophe , Random Fo es , and LSTM—compa ed o adi ional
app oaches—sugges s a p omising a enue o u he esea ch. Addi-
ionally, decision ee-based models ha e ecei ed limi ed a en ion in
he con ex o con aine eigh a e o ecas ing, despi e demons a ed
success in o he ields. The e o e, he p esen s udy seeks o add ess
hese gaps by sys ema ically compa ing he pe o mance o P ophe ,
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
100
Decision T ee, Random Fo es , and LSTM models. This compa a i e
analysis aims o elucida e he s eng hs and limi a ions o each app oach
and o iden i y he mos e ec i e me hodology o cap u ing he com-
plex, non-linea dynamics inhe en in con aine eigh a e da a.
3. Me hodology
3.1. Da a
Ocean eigh a es a e in luenced by a ious ac o s, including
supply and demand dynamics, ca go weigh and olume, and he dis-
ance o des ina ion (Khan & Hussain, 2022). Recen geopoli ical
de elopmen s—such as he Red Sea c isis, which p omp ed con aine
ships o bypass he Suez Canal in a o o he Sou h A ican Cape o Good
Hope—ha e u he impac ed hese a es. His o ical e en s, including
he global inancial c isis, he COVID-19 pandemic, and episodes o
o e capaci y, ha e also demons a ed signi ican e ec s on ocean eigh
a es (Naima e al., 2023; Wang e al., 2024).
To cap u e hese dynamics, his s udy u ilizes a iables ep esen ing
con aine shipping olume, con aine capaci y, and key economic in-
dica o s. Speci ically, Asia-Eu ope capaci y and Asia-No h Ame ica
capaci y a e employed o quan i y shipping capaci y, while Asia-Eu ope
shipping olume and Asia-No h Ame ica shipping olume measu e
shipping olume. Ocean eigh a es a e examined along he No h
Ame ican, Eu opean, and Medi e anean Sea ou es, as de i ed om he
Shanghai Con aine ized F eigh Index (SCFI). The economic indica o s
inco po a ed in o he analysis include Global Economic Policy Unce -
ain y (EPU), he G20 Composi e Leading Indica o (CLI), and Global
Geopoli ical Risk (GPR). EPU is calcula ed based on he equency o
e ms such as "economy" and "policy" in media a icles (Bake e al.,
2016), CLI e lec s expec a ions o u u e economic ac i i y (OECD), and
GPR quan i ies geopoli ical ensions based on news equency (Calda a
& Iaco iello, 2022).
This s udy u ilizes mon hly da a spanning om Janua y 2014 o June
2024. Due o he he e ogeneous uni s o measu emen (e.g., TEU e sus
USD/TEU), all ea u es a e no malized using a Min-Max scaling ech-
nique, which scales alues o he [0,1] ange. Subsequen ly, he
Johansen co-in eg a ion es is pe o med o asce ain he exis ence o
long- e m ela ionships among he a iables. The p esence o co-
in eg a ion jus i ies he use o he aw da a in he o ecas ing models.
Table 1 p o ides a summa y o he desc ip i e s a is ics o he da a used
in his s udy.
3.2. Fo ecas ing models
3.2.1. P ophe model
The P ophe model, in oduced by Taylo and Le ham (2018), is
speci ically designed o obus ime-se ies o ecas ing. I s p incipal
s eng h lies in i s capaci y o model seasonal pa e ns, long- e m ends,
and holiday e ec s independen ly, allowing o easy cus omiza ion o
speci ic business con ex s. This lexibili y enables P ophe o e ec i ely
cap u e he non-linea cha ac e is ics and pe iodic luc ua ions inhe en
in ocean eigh a e da a.
3.2.2. Decision ee and andom o es models
The Decision T ee model is a well-es ablished me hod in p edic i e
analy ics, o iginally concep ualized by Belson (1959) and u he
de eloped h ough he CART me hodology by B eiman e al. (1986).
Al hough decision ees a e in ui i e and easy o in e p e , hey a e
suscep ible o o e i ing. To add ess his limi a ion, Random Fo es —an
ensemble lea ning echnique p oposed by B eiman (2001)—agg ega es
he p edic ions o mul iple decision ees h ough a o ing mechanism.
This ensemble app oach mi iga es o e i ing and enhances he model’s
pe o mance, pa icula ly when dealing wi h high-dimensional da a.
3.2.3. LSTM model
Long Sho -Te m Memo y (LSTM) ne wo ks, in oduced by
Hoch ei e (1997), a e a specialized ype o ecu en neu al ne wo k
designed o o e come he limi a ions o adi ional RNNs in cap u ing
long- e m dependencies. LSTM ne wo ks inco po a e ga ing mecha-
nisms—namely, he inpu , o ge , and ou pu ga es—along wi h a
memo y cell ha p ese es and upda es in o ma ion as needed. These
ea u es add ess he anishing g adien p oblem, he eby enabling he
e icien lea ning o empo al dependencies and making LSTM pa icu-
la ly well-sui ed o ime-se ies o ecas ing asks such as p edic ing
ocean eigh a es.
3.2.4. Model se ings
Random Sea ch was employed o iden i y he op imal hype -
pa ame e con igu a ions o he Decision T ee, Random Fo es , and
P ophe models. In con as , he LSTM model was ained using a
consis en hype pa ame e con igu a ion ac oss all sec ions, a design
choice made o balance pe o mance and compu a ional demands.
Table 2 p esen s he sec ion-speci ic se ings used o hype pa ame e
op imiza ion.
3.2.5. Model alida ion
To assess he o ecas ing pe o mance o he models, his s udy em-
ploys a comp ehensi e se o e alua ion me ics: Mean Absolu e E o
(MAE), Mean Absolu e Pe cen age E o (MAPE), Mean Squa ed E o
(MSE), Roo Mean Squa ed E o (RMSE), No malized Mean Squa ed
E o (NMSE) and Symme ic Mean Absolu e Pe cen age E o (SMAPE).
MAE p o ides a di ec measu e o he a e age absolu e e o be ween
he p edic ed and ac ual alues. MAPE exp esses his e o as a
Table 1
Da a desc ip i e s a is ics.
Coun Mean S d Min Max 50 %
EPU* 126 208.86 72.57 86.63 431.73 208.31
GPR* * 126 105.78 33.75 58.42 318.95 102.19
CLI* ** 126 99.73 1.45 89.48 101.41 99.99
Capaci y_EUR* ** * 126 1858,579 153,128.7 1559,476 2142,365 1860,088
Capaci y_NA* ** * 126 2015,801 424,777.4 1462,072 2912,504 1876,666
Volume_EUR* ** * 126 1330,787 153,564.1 67,620 1616,700 1364,350
Volume_NA* ** * 126 1615,153 284,394.8 81,270 2119,400 1582,550
Med* ** ** 126 2019.58 2064.47 220.50 7522.75 984.71
NEUR* ** ** 126 1847.21 2055.13 223.50 7784.25 912.63
USWC* ** ** 126 2587.77 1895.83 796.50 8079.00 1800.13
USEC* ** ** 126 4091.98 2631.86 1589.50 2962.75 11,778.50
* Bake , S. R., Bloom, N., & Da is, S. J. (2016). "Measu ing Economic Policy Unce ain y." A ailable a : Policy Unce ain y. / Uni less Index
* * Calda a, D and Iaco iello, M. (2022). "Global Policy Unce ain y." A ailable a : Ma eo Iaco iello’s Websi e. / Uni less Index
* **OECD. "Composi e Leading Indica o (CLI) - G20." A ailable a : OECD. / Uni : Long- e m a e age =100
* ** *Bloombe g L.P. "Shipping Capaci y and Shipping Volume Da a." Bloombe g Te minal. / Uni : TEU
* ** **Shanghai Shipping Exchange. "Shanghai Con aine ized F eigh Index." A ailable a : Shanghai Shipping Exchange. / Uni : USD/TEU, Uni : USD/FEU
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
101
pe cen age, acili a ing compa isons ac oss models wi h di e en scales.
MSE and RMSE, by emphasizing la ge e o s h ough squa ing, o e
insigh s in o he models’ sensi i i y o signi ican de ia ions. NMSE
no malizes he e o ela i e o he a iance o he da a, allowing o a
ela i e assessmen o p edic i e pe o mance. MAPE has a p oblem ha
he e o ises in ini ely when he ac ual alue is close o 0, and SMAPE
is used o compensa e o his. SMAPE e lec s he symme y be ween
he p edic ed and ac ual alues and wo ks s ably nea ze o. Lowe alues
ac oss hese me ics indica e be e model pe o mance, guiding he
selec ion o he mos e ec i e o ecas ing app oach o con aine eigh
a es.
4. Resul s
4.1. Johansen es
The Johansen co-in eg a ion es was conduc ed o asce ain he
exis ence o a long- e m equilib ium ela ionship among he a iables.
In his es , he null hypo hesis posi s ha he numbe o co-in eg a ing
ec o s is less han a speci ied ank (n < ), whe eas he al e na i e
hypo hesis asse s ha is less han o equal o n. I he es s a is ic
exceeds he c i ical alue, he null hypo hesis is no ejec ed; con e sely,
i is ejec ed when he es s a is ic alls below he c i ical alue.
As summa ized in Table 3, he es esul s con i m he p esence o co-
in eg a ion ac oss all ou es examined. Consequen ly, despi e any non-
s a iona i y in he indi idual se ies, he exis ence o a long- e m ela-
ionship jus i ies he applica ion o he aw da a in subsequen modeling.
Speci ically, he numbe o co-in eg a ions was de e mined o lie wi hin
ange 2 < ≤3 o bo h he USEC and USWC ou es, and wi hin 1 < ≤2
o he MED and NEUR ou es.
4.2. Model se ing
All models u ilizing no malized da a we e ained using 80 % o he
da ase , wi h he emaining 20 % ese ed o es ing. The pe o mance
e alua ions p esen ed in Table 5 a e based on models ained using he
hype pa ame e con igu a ions de ailed in Table 4. Speci ically, Table 4
Table 2
Sec ion se ings o hype pa ame e op imiza ion.
Decision T ee Random Fo es P ophe LSTM
Max Dep h: 1–20 N Es ima o s: 1–100
Max Dep h: 1–20
Min Samples Spli : 10–100
Min Samples Lea : 1–20
Max Fea u es: 1–15
Max Lea Nodes: 1–100
Min Weigh F ac ion Lea : 0–1
G ow h: Linea , Logis ic
Daily Seasonali y: T ue, False
Weekly Seasonali y: T ue, False
Yea ly Seasonali y: T ue, False
Changepoin P io Scale: 0.001–100(Log Scale)
Seasonali y P io Scale: 0.001–100(Log Scale)
Sequence Leng h: 3
D opou : 0.5
Uni s: 128
Epochs: 100
Ba ch Size: 64
Op imize : Adam
Lea ning Ra e: 0.001
Loss: MSE
Min Samples Spli : 10–100
Min Samples Lea : 1–20
Max Fea u es: 1–15
Max Lea Nodes: 1–100
Min Weigh F ac ion Lea : 0–1
C i e ion: squa ed_e o , iedman_mse, absolu e_e o , poisson
Table 3
Johansen es .
Rou e _0 _1 Tes S a is ic C i ical Value (%)
USEC 0 6 125.2 95.75
1 6 84.97 69.82
2 6 51.54 47.85
3 6 28.44 29.80
USWC 0 6 134.0 95.75
1 6 87.06 69.82
2 6 53.06 47.85
3 6 27.28 29.80
MED 0 6 125.5 95.75
1 6 75.16 69.82
2 6 44.02 47.85
NEUR 0 6 129.0 95.75
1 6 76.92 69.85
2 6 45.93 47.85
Table 4
Se ings o hype pa ame e .
Rou e Model Hype pa ame e
USWC Decision
T ee
andom_s a e=42, c i e ion=’squa ed_e o ’,
max_dep h=19, max_ ea u es=2, max_lea _nodes=68,
min_samples_lea =15, min_samples_spli =2,
min_weigh _ ac ion_lea =0
Random
Fo es
andom_s a e=42, max_dep h=6, max_ ea u es=7,
max_lea _nodes=81, min_samples_lea =1,
min_samples_spli =16, min_weigh _ ac ion_lea =0,
n_es ima o s=21, boo s ap=T ue, oob_sco e=False,
n_jobs=None
LSTM Sequence Leng h=3, D opou =0.5, Uni s=128,
Epochs=100, Ba ch Size=64, Op imize =Adam, Lea ning
Ra e=0.001, Loss=MSE
P ophe g ow h=’logis ic’, changepoin _p io _scale=0.0081,
seasonali y_p io _scale=572.237,
yea ly_seasonali y=T ue, daily_seasonali y=False,
weekly_seasonali y=False
USEC Decision
T ee
andom_s a e=42, c i e ion=’squa ed_e o ’,
max_dep h=15, max_ ea u es=13, max_lea _nodes=27,
min_samples_lea =13, min_samples_spli =52,
min_weigh _ ac ion_lea =0
Random
Fo es
andom_s a e=42, max_dep h=6, max_ ea u es=3,
max_lea _nodes=96, min_samples_lea =11,
min_samples_spli =21, min_weigh _ ac ion_lea =0,
n_es ima o s=26, boo s ap=T ue, oob_sco e=False,
n_jobs=None
LSTM Sequence Leng h=3, D opou =0.5, Uni s=128,
Epochs=100, Ba ch Size=64, Op imize =Adam, Lea ning
Ra e=0.001, Loss=MSE
P ophe g ow h=’logis ic’, changepoin _p io _scale=0.0035,
seasonali y_p io _scale=0.0013, yea ly_seasonali y=T ue,
daily_seasonali y=False, weekly_seasonali y=False
MED Decision
T ee
andom_s a e=42, c i e ion=’squa ed_e o ’,
max_dep h=19, max_ ea u es=9, max_lea _nodes=74,
min_samples_lea =6, min_samples_spli =16,
min_weigh _ ac ion_lea =0
Random
Fo es
andom_s a e=42, max_dep h=7, max_ ea u es=14,
max_lea _nodes=38, min_samples_lea =10,
min_samples_spli =18, min_weigh _ ac ion_lea =0,
n_es ima o s=39, boo s ap=T ue, oob_sco e=False,
n_jobs=None
LSTM Sequence Leng h=3, D opou =0.5, Uni s=128,
Epochs=100, Ba ch Size=64, Op imize =Adam, Lea ning
Ra e=0.001, Loss=MSE
P ophe g ow h=’logis ic’, changepoin _p io _scale=0.0081,
seasonali y_p io _scale=0.6136, yea ly_seasonali y=T ue,
daily_seasonali y=False, weekly_seasonali y=False
NEUR Decision
T ee
andom_s a e=42, c i e ion=’squa ed_e o ’,
max_dep h=19, max_ ea u es=3, max_lea _nodes=70,
min_samples_lea =3, min_samples_spli =27,
min_weigh _ ac ion_lea =0
Random
Fo es
andom_s a e=42, max_dep h=13, max_ ea u es=12,
max_lea _nodes=76, min_samples_lea =2,
min_samples_spli =40, min_weigh _ ac ion_lea =0,
n_es ima o s=57, boo s ap=T ue, oob_sco e=False,
n_jobs=None
LSTM Sequence Leng h=3, D opou =0.5, Uni s=128,
Epochs=100, Ba ch Size=64, Op imize =Adam, Lea ning
Ra e=0.001, Loss=MSE
P ophe g ow h=’logis ic’, changepoin _p io _scale=0.0023,
seasonali y_p io _scale=0.0327, yea ly_seasonali y=T ue,
daily_seasonali y=False, weekly_seasonali y=False
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
102
ou lines he op imized hype pa ame e se ings o each o he ou
models, while Table 5 compa es he p edic i e pe o mance o hese
models ollowing aining wi h he speci ied con igu a ions.
4.3. Model compa ison
The p edic i e pe o mance o he ou o ecas ing models—Decision
T ee, Random Fo es , LSTM, and P ophe —was igo ously e alua ed
using six me ics: Mean Squa ed E o (MSE), Roo Mean Squa ed E o
(RMSE), No malized Mean Squa ed E o (NMSE), Mean Absolu e E o
(MAE), Mean Absolu e Pe cen age E o (MAPE) and Symme ic Mean
Absolu e Pe cen age E o (SMAPE). Table 5 p esen s a de ailed sum-
ma y o hese e alua ion indica o s o each shipping ou e.
O e all, he Decision T ee and Random Fo es models demons a ed
supe io pe o mance compa ed o he LSTM and P ophe models when
assessed using absolu e e o me ics (MSE, RMSE, NMSE, and MAE).
This obse a ion sugges s ha ee-based me hods a e mo e e ec i e in
cap u ing he unde lying non-linea dynamics o con aine eigh a e
da a. The lowe e o alues achie ed by hese models indica e hei
enhanced capabili y o model he complex ela ionships p esen in he
da ase .
Fo he USWC ou e, he Random Fo es model achie ed he lowes
absolu e e o alues, eco ding an MSE o 0.0322, an RMSE o 0.1796,
an NMSE o 0.5853, and an MAE o 0.1007. Speci ically, he MSE ob-
ained by he Random Fo es model was app oxima ely 58.6 % lowe
han ha o he Decision T ee, 85.2 % lowe han ha o he P ophe
model, and 99.15 % lowe han ha o he LSTM model. These im-
p o emen s in e o me ics unde sco e he obus pe o mance o he
Random Fo es app oach in cap u ing he dynamics o he USWC ou e.
In con as , o he USEC ou e, he Decision T ee model exhibi ed
he mos a o able pe o mance. The Decision T ee eco ded an MSE o
0.0310, an NMSE o 0.4559, an RMSE o 0.1579, and an MAE o 0.1007,
ou pe o ming he Random Fo es model by a modes ma gin. Speci -
ically, i s MSE was abou 2.9 % lowe han ha o he Random Fo es ,
app oxima ely 87.2 % lowe han ha o he P ophe model, and nea ly
99.2 % lowe han ha o he LSTM model. This esul indica es ha , o
he USEC ou e, he Decision T ee model mo e e ec i ely cap u es he
unde lying da a s uc u e.
Fo he MED ou e, while he Random Fo es model gene ally ach-
ie ed lowe e o me ics—wi h an MSE o 0.0347, an RMSE o 0.1863,
and an NMSE o 0.4359— he Decision T ee model yielded he lowes
MAE a 0.1181. The pe o mance di e ences be ween hese models,
al hough small, sugges ha each model may ha e s eng hs in cap u ing
di e en aspec s o he da a a iabili y o he MED ou e.
On he NEUR ou e, he Random Fo es model again eco ded he
lowes absolu e e o alues, wi h an MSE o 0.0447, an RMSE o 0.2115,
an NMSE o 0.5434, and an MAE o 0.1297. The Random Fo es ’s pe -
o mance on his ou e was app oxima ely 29.3 % be e in e ms o
MSE, 15.9 % be e in RMSE, and 30.5 % be e in NMSE compa ed o
he Decision T ee model. None heless, he di e ences in pe o mance
me ics be ween he wo models on he NEUR ou e we e ela i ely
ma ginal.
A compa ison be ween MAPE and SMAPE e eals ha SMAPE
consis en ly demons a es be e pe o mance ac oss all ou es. This
sugges s ha many o he ac ual alues used in he MAPE denomina o
a e close o ze o, he eby in la ing he MAPE e o . To add ess his issue,
bo h MAPE and SMAPE we e join ly examined, and models exhibi ing
ela i ely s able alues ac oss bo h me ics we e selec ed. Based on his
comp ehensi e e alua ion, he Decision T ee and Random Fo es models
gene ally ou pe o med he LSTM and P ophe models ac oss mos
ou es.
Despi e he a o able ou comes o bo h he Decision T ee and
Random Fo es models wi h espec o absolu e e o me ics, he ela-
i e e o measu e—MAPE and SMAPE— e ealed conside ably highe
p edic ion e o a ios ac oss all models. This ele a ed MAPE and
SMAPE is la gely a ibu able o he inc eased ola ili y in con aine
eigh a es ollowing dis up ions such as he COVID-19 pandemic and
he Red Sea c isis. Ele a ed MAPE and SMAPE alues ac oss all models
a e la gely a ibu able o he inc eased ola ili y in con aine eigh
a es ollowing dis up ions such as he onse o he COVID-19 pandemic
and he Red Sea c isis, which ampli ied global supply chain dis u bances
and esul ed in la ge pe cen age e o s, as illus a ed in Fig. 1.
Compa ing SMAPE, i was analyzed ha USEC and MED ha e good
decision ee pe o mance, and USWC and NEUR ha e good Random
Fo es pe o mance. No ably, a compa ison o he MAPE alues ac oss
he ou es indica es ha he Decision T ee model consis en ly exhibi s
lowe ela i e e o a ios han he Random Fo es model. Speci ically,
he Decision T ee model achie ed MAPE imp o emen s o app oxi-
ma ely 91.8 % o he USWC ou e, 52.1 % o he USEC ou e, 43.5 %
o he MED ou e, and 22.7 % o he NEUR ou e when compa ed o he
Random Fo es model. The e o e, bo h MAPE and SMAPE adop ed a
ela i ely s able Decision T ee.
Finally, using he Decision T ee model, he in luence o a ious
ea u es on con aine eigh a es was isualized ac oss all ou es. SHAP
(SHapley Addi i e Explana ions) was employed o quan i a i ely assess
he impac o hese ea u es on he machine lea ning p edic ion ou -
comes. The g aphical ep esen a ions o ea u e in luence o he
Random Fo es , P ophe , and LSTM model p edic ions a e p o ided in
he Appendix. Speci ically, bo h he Decision T ee and Random Fo es
models u ilized SHAP o ea u e impac isualiza ion, he P ophe
model e alua ed ea u e in luence h ough eg ession coe icien s, and
he LSTM model illus a ed in luence based on he a e age weigh s o i s
ea u e se .
5. Discussion
The empi ical analysis p esen ed in Figs. 2–5elucida es he mul i-
ace ed de e minan s o con aine eigh a es, emphasizing he c i ical
oles o bo h supply-side and demand-side ac o s. In pa icula , capaci y
and olume exe subs an ial in luences on eigh a e luc ua ions
ac oss all examined ou es. These indings ein o ce he heo e ical
amewo k ad anced in ea lie s udies (Jeon e al., 2020; Sca si, 2007),
which posi ha shipping capaci y and shipping olume a e p ima y
d i e s o eigh a e dynamics.
Beyond hese undamen al supply-demand a iables, ou analysis
highligh s he pi o al ole o he G20 CLI on ou es o he han USEC. As a
Table 5
E alua ion indica o s o models.
Rou e indica o Decision
T ee
Random
Fo es
LSTM P ophe
USWC MSE 0.0778 0.0322 3788.1917 0.2171
RMSE 0.2790 0.1796 61.5482 0.4659
NMSE 0.8638 0.5853 75024.7661 1.7127
MAE 0.1615 0.1007 54.7361 0.4194
MAPE 36.9699 448.6831 39944.7295 330.4461
SMAPE 50.0792 47.3774 83.4179 73.5523
USEC MSE 0.0249 0.0310 3873.1899 0.2420
RMSE 0.1579 0.1761 62.2349 0.4920
NMSE 0.3618 0.4559 56941.6229 2.3690
MAE 0.1007 0.1146 55.1988 0.4235
MAPE 53.8369 112.3587 87888.3997 365.6392
SMAPE 42.1499 55.9611 76.8371 88.9091
MED MSE 0.0427 0.0347 4025.4914 0.1472
RMSE 0.2068 0.1863 63.4467 0.3837
NMSE 0.5136 0.4359 136494.9406 1.6701
MAE 0.1181 0.1265 56.1486 0.3487
MAPE 45.8080 81.1313 23215.9150 152.6911
SMAPE 43.2302 55.3836 41.5510 71.0931
NEUR MSE 0.0632 0.0447 3363.9074 0.1842
RMSE 0.2514 0.2115 57.9992 0.4292
NMSE 0.9562 0.5434 40844.9406 1.9608
MAE 0.1559 0.1297 52.3133 0.3877
MAPE 81.4221 105.2359 38170.2764 342.0722
SMAPE 66.1932 60.7961 68.7988 93.0079
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
103
comp ehensi e measu e ha encapsula es key ace s o economic ac i-
i y—such as manu ac u ing pe o mance, consume con idence, and
new o de le els— he CLI se es as a obus p oxy o assessing he
o e all economic en i onmen . High CLI alues ypically signal an
expanding economy, p omp ing shipping companies o inc ease capac-
i y in an icipa ion o ising demand. Con e sely, low CLI eadings sug-
ges economic s agna ion o con ac ion, he eby incen i izing
s a egies such as capaci y educ ion o blank sailing. The s ong co -
ela ion obse ed be ween CLI luc ua ions and eigh a e mo emen s
unde sco es i s u ili y as a leading indica o o ma i ime anspo
planning.
The analysis also e eals ha EPU signi ican ly impac s con aine
eigh a es on mos ou es, wi h he no able excep ion o he NEUR
ou e. In ecen yea s, ongoing a i dispu es and an i-dumping in es-
iga ions—pa icula ly be ween Eu ope and he Uni ed S a es
Fig. 1. SCFI.
Fig. 2. The in luence o ea u e (USWC).
Fig. 3. The in luence o ea u e (USEC).
Fig. 4. The in luence o ea u e (MED).
Fig. 5. The in luence o ea u e (NEUR).
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
104
conce ning impo s om China—ha e ele a ed EPU le els. These con-
di ions ha e led shippe s o inc ease impo olumes p eemp i ely,
which, in u n, ha e d i en shipping companies o expedi e he imple-
men a ion o Peak Season Su cha ges (PSS). A case in poin is Mae sk’s
decision o en o ce PSS ea lie han usual o shipmen s om he Asia
Paci ic egion o No h Ame ica and Canada, a esponse a ibu ed o
su ging impo olumes obse ed in ea ly July 2024. The mu ed in lu-
ence o EPU on he NEUR ou e is likely due o he s abilizing e ec o
he Eu opean Union’s in eg a ed economic amewo k, which bu e s
agains policy-induced ola ili y.
Fu he mo e, he GPR indica o has demons a ed a p onounced
impac on he MED and NEUR ou es. The ecen Red Sea c isis, cha -
ac e ized by he occupa ion o he Red Sea by Yemeni Hou hi ebels
s a ing in la e 2023, led o signi ican dis up ions in Suez Canal ope -
a ions. This geopoli ical ins abili y o ced shipping companies o e ou e
essels ia he longe Cape o Good Hope, he eby inc easing ansi
dis ances and ope a ional cos s. Quan i a i e analysis, as depic ed in
Fig. 6, indica es ha du ing he i s hal o 2024, con aine eigh a es
on he MED ou e inc eased by app oxima ely 130 % ela i e o he
same pe iod in 2023, while he NEUR ou e expe ienced an inc ease o
abou 226 %. These indings highligh he c i ical impac o exogenous
geopoli ical shocks on eigh a e ola ili y.
The discussion unde sco es ha con aine eigh a e de e mina ion
is a complex in e play o inhe en supply-demand dynamics and exog-
enous ac o s, including mac oeconomic indica o s and geopoli ical
isks. The in eg a ion o hese di e se a iables in o ou o ecas ing
models no only enhances p edic i e pe o mance bu also o e s sig-
ni ican insigh s o s a egic decision-making in he ma i ime anspo
sec o . Fu u e esea ch should u he explo e hese in e dependencies,
pa icula ly unde condi ions o heigh ened ma ke ola ili y and
geopoli ical unce ain y, o de elop mo e esilien and adap i e o e-
cas ing amewo ks.
6. Conclusion
Con aine eigh a es play a pi o al ole in logis ics cos manage-
men and s a egic decision-making wi hin he shipping indus y. The
cyclical na u e o ship o de ing—whe e high eigh a es p omp new
o de s ha a e deli e ed wo o h ee yea s la e , po en ially leading o
o e capaci y and subsequen a e declines—exace ba es he inhe en
ma ke ola ili y. Mo eo e , he inc easing egula o y p essu es o
educe ca bon emissions ha e d i en shipping companies o in es in
eco- iendly essels. Consequen ly, accu a e o ecas ing o ocean eigh
a es becomes essen ial o e ec i e isk managemen in an indus y
cha ac e ized by long- e m planning and signi ican capi al in es men s.
This s udy unde ook a compa a i e analysis o se e al o ecas ing
models, including Decision T ee, Random Fo es , LSTM, and P ophe , o
de e mine hei ela i e pe o mance in p edic ing con aine eigh
a es. Ou empi ical esul s indica e ha , in e ms o absolu e e o
me ics, bo h he Decision T ee and Random Fo es models exhibi
s ong p edic i e capabili ies. In his s udy, he model demons a ing he
bes o e all pe o mance was selec ed based on wo e alua ion ap-
p oaches: (1) compa ison o MSE, RMSE, NMSE, and MAE, and (2)
compa ison o MAPE and SMAPE. In he i s e alua ion, bo h he
Random Fo es and Decision T ee models exhibi ed s ong pe o mance,
wi h Random Fo es ou pe o ming in some me ics. Howe e , in he
second e alua ion using MAPE and SMAPE, he Decision T ee model was
ul ima ely selec ed due o i s ela i ely s able and supe io pe o mance
ac oss bo h indica o s. These esul s highligh ha he op imal model
may a y depending on he chosen e alua ion me ic. No ably, he De-
cision T ee model eme ged as he mos e ec i e when assessed using he
ela i e e o me ic, MAPE and SMAPE, despi e o e all MAPE and
SMAPE pe o mance emaining subop imal ac oss all models. The
ele a ed MAPE and SMAPE alues can be a ibu ed o he ampli ied
impac o ex e nal shocks, pa icula ly EPU and GPR, which ha e
inc easingly in luenced ma ke dynamics. This con i ma ion o he
impac o EPU and GPR is expec ed o help he shipping indus y
s akeholde s make decisions.
The indings o his s udy unde sco e he c i ical in luence o bo h
adi ional supply-demand ac o s and ex e nal economic and geopo-
li ical shocks on eigh a e o ecas ing. Howe e , he pe sis ence o
ela i ely high p edic ion e o a ios highligh s a signi ican limi a ion
o he cu en modeling app oaches. This sugges s he need o he
de elopmen o mo e sophis ica ed hyb id models—po en ially in e-
g a ing app oaches such as ANN and RNN— ha a e be e equipped o
cap u e he complex, non-linea in e ac ions be ween exogenous ac o s
Fig. 6. MED SCFI and NEUR SCFI (2023.01 – 2023.06 / 2024.01 – 2024.06).
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
105
and eigh a e dynamics.
Fu u e esea ch should ocus on enhancing p edic i e accu acy by
inco po a ing addi ional ex e nal a iables and explo ing ensemble
me hods ha syne gis ically combine he s eng hs o a ious machine
lea ning and deep lea ning models. Such ad ancemen s will be c ucial
o de eloping obus o ecas ing ools capable o suppo ing s a egic
decision-making in an inc easingly ola ile global shipping
en i onmen .
CRediT au ho ship con ibu ion s a emen
Cha Junhee: Da a cu a ion, Fo mal analysis. Kim Namhun:
Concep ualiza ion, Da a cu a ion. Jeon Junwoo: Concep ualiza ion,
Fo mal analysis, Supe ision.
Decla a ion o Compe ing In e es
The au ho s decla e ha he e a e no con lic s o in e es ega ding
he publica ion o his pape .
Appendix
Fig. 7. The In luence o Fea u e (USWC), Random Fo es
Fig. 8. The In luence o Fea u e (USEC), Random Fo es
Fig. 9. The In luence o Fea u e (MED), Random Fo es
N. Kim e al.
The Asian Jou nal o Shipping and Logis ics 41 (2025) 99–109
106
