What microeconomic fundamentals drove global oil prices during 1986-2020?

Mallia is, Anas asios G.; Mallia is, Ma y E. A icle Wha mic oeconomic undamen als d o e global oil p ices du ing 1986-2020? Jou nal o Risk and Financial Managemen P o ided in Coope a ion wi h: MDPI – Mul idisciplina y Digi al Publishing Ins i u e, Basel Sugges ed Ci a ion: Mallia is, Anas asios G.; Mallia is, Ma y E. (2021) : Wha mic oeconomic undamen als d o e global oil p ices du ing 1986-2020?, Jou nal o Risk and Financial Managemen , ISSN 1911-8074, MDPI, Basel, Vol. 14, Iss. 8, pp. 1-13, h ps://doi.o g/10.3390/j m14080391 This Ve sion is a ailable a : h ps://hdl.handle.ne /10419/258495 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/ Jou nal o Risk and Financial Managemen A icle Wha Mic oeconomic Fundamen als D o e Global Oil P ices du ing 1986–2020? Anas asios G. Mallia is 1,* and Ma y Mallia is 2   Ci a ion: Mallia is, Anas asios G., and Ma y Mallia is. 2021. Wha Mic oeconomic Fundamen als D o e Global Oil P ices du ing 1986–2020? Jou nal o Risk and Financial Managemen 14: 391. h ps://doi. o g/10.3390/j m14080391 Academic Edi o : Robe Hudson Recei ed: 6 July 2021 Accep ed: 19 Augus 2021 Published: 21 Augus 2021 Publishe ’s No e: MDPI s ays neu al wi h ega d o ju isdic ional claims in published maps and ins i u ional a il- ia ions. Copy igh : © 2021 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h ps:// c ea i ecommons.o g/licenses/by/ 4.0/). 1 Economics Depa men and Finance Depa men , The Quinlan School o Business, Loyola Uni e si y Chicago, 16 E. Pea son, Chicago, IL 60611, USA 2In o ma ion Sys ems & Supply Chain Managemen Depa men , The Quinlan School o Business, Loyola Uni e si y Chicago, 16 E. Pea son, Chicago, IL 60611, USA; [email p o ec ed] *Co espondence: [email p o ec ed] Abs ac : The global inancial c isis o 2007–2009 caused majo economic dis u bances in he oil ma ke . In his pape , we conside i e a iables ha desc ibe he mic oeconomics o he supply o and demand o oil, and e alua e hei impo ance be o e, du ing and a e he global inancial c isis. We conside i e dissimila egimes du ing he pe iod o Janua y 1986 o he end o 2020: wo egimes p io o he global inancial c isis, he egime du ing he c isis, and wo egimes a e he c isis. The main hypo hesis es ed is ha oil undamen als o supply and demand emained impo an , e en hough he i e egimes we e dissimila . We buil i e boos ed and o e - i ed neu al ne wo ks o cap u e he exac ela ionships be ween spo oil p ices and oil da a ela ed o hese p ices. This analysis shows ha , while he inpu s in o an accu a e neu al ne wo k can emain he same, he impac o each a iable can change conside ably du ing di e en egimes. Keywo ds: oil p ice egimes; neu al ne wo k me hodology; 2007–2009 global inancial c isis; s uc- u al b eaks 1. In oduc ion Global oil ma ke s a e complex, and he p ice o oil exhibi s signi ican ola ili y. Du ing he pas ou decades, he p ice o oil was as low as abou USD 20 in he la e 1990s and as high as USD 150 be o e he global inancial c isis. Baumeis e and Kilian (2016) conduc ed an excellen o e iew o he beha io o oil p ices o e he pas o y yea s, wi h a special ocus on he luc ua ions o hese p ices. They emphasized ha luc ua ions o such magni ude con ain a la ge componen o su p ise, in he sense ha o dina y economic o esigh could no pene a e he dynamic succession o egime shi ing in oil ma ke s. Fo example, e en s such as he 1973–1974 oil c isis, which saw delibe a e cu s in oil p oduc ion in A ab OPEC coun ies, we e ollowed by he unexpec ed pe iod o G ea Mode a ion du ing he mid-1980s. The mode a ion o in la ion and he educ ion o eal GDP luc ua ions c ea ed con- di ions o inc eased isk- aking and he housing bubble o 2002 o 2007, which was cha ac e ized by an associa ed global boom in eme ging ma ke s such as China, India and B azil ha ueled big inc eases in he p ice o oil. Then, he wo ld expe ienced he un o eseen global inancial c isis (GFC) and he ushe ing in o uncon en ional mone a y policies a ound he wo ld, which b ough oil p ices om abou USD 150 down o USD 40. The pos -c isis pe iod du ing 2009 o 2013 eco ded s eady inc eases in he p ice o oil, om a low o abou USD 40 o a high o USD 110, p ima ily because o he economic eco e y o he global economy. The mos ecen pe iod, om June 2013 o Decembe 2020, shows how U.S. acking echnology, wi h i s s ong shale p oduc ion, combined wi h ising OPEC ou pu and slowing demand om China, ha e con ibu ed o d ama ic declines in oil p ices, om a high o USD 100 in June 2014 o a low o USD 30 in Feb ua y 2016. Du ing he 2016–2020 pe iod, oil p ices luc ua ed be ween a low o USD 30 in Feb ua y 2016 o J. Risk Financial Manag. 2021,14, 391. h ps://doi.o g/10.3390/j m14080391 h ps://www.mdpi.com/jou nal/j m J. Risk Financial Manag. 2021,14, 391 2 o 13 a high o USD 75 in Sep embe 2018, o d op again o USD 20, one o oil’s lowes p ices, du ing Ap il 2020, when he COVID-19 c isis des abilized he en i e global economy. Baumeis e and Kilian (2014,2015,2016) ca ied ou a comp ehensi e analysis o his pe iod, while o he au ho s ha e examined speci ic opics and pe iods, such as Hamil on (1983); Be nanke e al. (1997); Nako and Pesca o i (2010); Kilian (2008). The ocus o his pape is he p icing o oil and i s de e minan s. We conside i e a iables ha desc ibe he mic oeconomics o he supply o , and demand o oil, mo i a ed by a simila app oach p oposed by Kilian (2009). We e alua e hei impo ance as oil de e minan s and hei a iabili y be o e, du ing, and a e he GFC. We conside i e dissimila egimes du ing he pe iod o Janua y 1986 o he end o 2017: wo egimes p io o he global inancial c isis, he egime du ing he c isis, and wo egimes a e he c isis. Mallia is and Mallia is (2020) s udied he beha io o he global p ice o oil using mac oeconomic a iables and employing he me hodology o o e lapping eg essions. Bha e al. (2021) ocused on U.S. oil p oduc ion. The classic pape on egime changes is Ang and Timme mann’s (2012), which ins uc s us abou he me hodology used. Economis s a e amilia wi h wo egime me hodologies om he analysis o business cycles. One egime con ains pe iods o economic g ow h and he nex egime includes pe iods o decline. The undamen al assump ion is ha all egimes o g ow h a e simila in all o he espec s; he same is assumed o egimes o decline. Howe e , egimes need no epea hemsel es be ween jus he wo s a es o g ow h and decline. Mos echniques conside wo egimes, and in he case o oil, hese ha e been applied ex ensi ely in pape s such as Bha and Mallia is’s (2011) and Tsai’s (2015). Rela ed pape s a e Nogue a’s (2013); Salisu and Fasanya’s (2013); Xiong e al.’s (2013). In ou case, we iden i y i e egimes, desc ibed in Sec ion 2, and assume ha each such egime is dis inc i e. The main hypo hesis es ed is ha he oil undamen als o supply and demand emain impo an e en hough he i e egimes a e dissimila . We build i e boos ed and o e - i ed neu al ne wo ks o cap u e he exac ela ionships be ween spo oil p ices and oil da a ela ed o hose p ices. O e i ing in ol es le ing he model ain on a speci ic da a se un il i i s ha da a e y closely. Such a model canno gene alize well o o he da ase s, bu can explain one speci ic se well. This analysis shows ha , while he inpu s in o an accu a e neu al ne wo k can emain he same, he impac o each a iable can change conside ably du ing di e en egimes. We ind ha he shi s in impac s o he a ious inpu s a e g ea enough o suppo he hypo hesis ha he e a e impo an s uc u al b eaks be ween pe iods. The e a e many easons o a oid o e i ing a neu al ne wo k, pa icula ly because i des oys any chance o building a gene alized model ha will be sui able o e ime. Howe e , in his pape , we use he downsides o o e i ing o ou bene i . By o e i ing models o i e di e en ime pe iods, we a e able o highligh why gene aliza ion is di icul o e he en i e pe iod. By building, hen compa ing, i e bespoke ne wo ks, we can iden i y he speci ic a eas whe e gene al models will need o imp o e in o de o o ecas well. Al hough we do no pe o m any o ecas ing, he neu al ne wo k me hodology applied in his pape , he mic oeconomic da a o he independen inpu s used, and he esul s ob ained p o ide an excellen impe us o u he wo k in his a ea along he lines o Chen’s (2014); Jammazi and Aloui’s (2012) and Shin e al.’s (2013). In his pape , we build i e neu al ne wo k models and o e i hem in o de o cap u e he ela ionships among he inpu s in he ways ha mos pe ec ly ma ch hei con igu a ions in a speci ic pe iod. As indica ed, i is no ou pu pose o build o ecas ing models wi h alida ion se s. Ra he , we wan o c ea e desc ip i e models ha encapsula e exis ing ela ionships in di e en egimes o he da a se . This will help us o unde s and shi s in a iable impac o e he a ious egimes. J. Risk Financial Manag. 2021,14, 391 3 o 13 2. Da a The da a o his s udy comp ise inpu s ela ed o oil and supply and demand o oil. The e a e i e a iables: he spo p ice o oil, he s o ed s ocks o oil excep o ha held by he go e nmen as s a egic pe oleum ese es, ne impo s o oil, oil p oduc ion, and p oduc supplied. The US Ene gy In o ma ion Adminis a ion (EIA) uses p oduc supplied as a ep esen a ion o he amoun o pe oleum p oduc s consumed. All da a a e mon hly, a he han seasonally adjus ed, and we e downloaded om he EIA. They co e he ime pe iod om Janua y 1986 h ough o he end o 2020. Table 1gi es he sou ce o he da a o each o hese a iables, and he abb e ia ed name used in he models. Table 1. Va iables and hei sou ces. Va iable Da a Sou ce Sho Name Sou ce URL P oduc ion o Oil in he U.S. MCRFPUS1 om EIA P od h ps://www.eia.go /dna /pe /pe _c d_c pdn_adc_ mbbl_m.h m (accessed on 14 Ma ch 2021) Ne Impo s o Oil in he U.S. MTTNTUS2 om EIA Ne Imp h ps://www.eia.go /dna /pe /his /Lea Handle . ashx?n=PET&s=MTTNTUS2& =M (accessed on 14 Ma ch 2021) Oil S ocks excluding he S a egic Pe oleum Rese e MTESTUS1 om EIA S ks h ps://www.eia.go /dna /pe /his /Lea Handle . ashx?n=PET&s=MTESTUS1& =M (accessed on 14 Ma ch 2021) P oduc Supplied (Consump ion) o Oil MTTUPUS1 om EIA Con h ps://www.eia.go /dna /pe /pe _cons_psup_dc_ nus_mbbl_m.h m (accessed on 14 Ma ch 2021) Spo P ice o Cushing Oil RWTC om EIA Spo P h ps://www.eia.go /dna /pe /pe _p i_sp _s1_m.h m (accessed on 14 Ma ch 2021) Mon h o he yea De i ed om da a abo e Mon h The spo p ice o oil is he p ice o a single ansac ion o deli e y o a de e mined quan i y o oil o a speci ied place. Demand o oil does no change apidly, bu is seasonal. As demand a ies, we see i s e ec s on p oduc ion, on impo s, and on he amoun o oil in s o age. The s ocks o oil, excluding he s a egic pe oleum ese es, includes he in en o ies s o ed o u u e use, and is epo ed in housands o ba els on he las day o he week. The s a egic pe oleum ese e is an amoun o s ock main ained by he go e nmen o use du ing any pe iod whe e he e is a majo dis up ion in supply. The alues o ne impo s a e measu ed in housands o ba els impo ed, and include oil om he 50 s a es, he Dis ic o Columbia, and U.S. possessions and e i o ies. Ne impo s is he amoun o impo s minus he amoun o expo s. P oduc ion is measu ed in housands o ba els. P oduc ion co e s he olume o oil p oduced om U.S. oil ese oi s. I includes he olume om he poin o cus ody ans e o ucks, pipelines, o o he means, wi h he in en o be anspo ed o e ine ies. The quan i ies a e es ima ed by he s a e and summed o he Pe oleum Adminis a ion dis ic (PADD) and hen o he U.S. le el. The P oduc Supplied is used as a s and-in o he consump ion o pe oleum p oduc s, since i is calcula ed as he amoun o pe oleum p oduc s emo ed om p incipal o igins. These i e a iables o m he undamen als o supply: p oduc ion, plus ne impo s, plus oil s ock (in en o y) and oil demand, which all de e mine he p ice o oil in he U.S. Va iables we e hen scaled o be be ween ze o and one. This se o a iables includes he abb e ia ion o he o iginal a iable name ollowed by Scld. Scaling allows he model o ea he a iables on a mo e equal basis, bu all alues will be posi i e, as hey e e o quan i ies and p ices. Scaling he da a o be be ween 0 and 1, e e ed o as Fea u e Scaling in machine lea ning (ML), is an impo an p ocess, especially when neu al ne wo ks a e being used. Because neu al ne wo k algo i hms use g adien descen as a p ocess, hey will ha e a be e and mo e apid con e gence o he a ge when ea u e scaling has been used. Wi h ML algo i hms ha calcula e dis ances, la ge alued a ibu es can domina e J. Risk Financial Manag. 2021,14, 391 4 o 13 he ou pu because he algo i hm is a ec ed by he ela i e sizes o he a iables. Thus, scaling he a iables speeds up he lea ning phase in a neu al ne wo k eed o wa d back p opaga ion algo i hm, and helps p e en a iables wi h la ge anges om o e -weigh ing node connec ions o a iables wi h smalle anges (Han e al. (2012); Guidici (2003); Weiss and Indu khya (1998)). The a ge o each ne wo k was he scaled alue o he spo p ice o oil. The da a se was spli in o i e pa s using he s uc u al b eaks c ea ed by he in e en- ing economic c isis. These i e ime pe iods a e shown in Table 2, along wi h he beginning and ending da es, and he numbe o mon hs in ol ed. Ou en i e da a se begins in Janua y o 1986 and shows a sideways mo emen o spo p ices un il la e 2001, ollowed by a s eady g ow h o spo p ices, wi h some noise, un il la e 2007. The c isis pe iod is he one de e mined by he Na ional Bu eau o Economic Resea ch and coincides wi h he G ea Recession ha began in Decembe 2007 and ended in May/June 2009. Du ing his pe iod, he oil bubble con inued o g ow in la e 2007 and ea ly 2008, and hen collapsed wi h Lehman B o he s’ bank up cy in mid-Sep embe 2008. A small bounce a he bo om led e en ually o a esump ion in he s eady g ow h o p ices a e mid-2009. A e his ime pe iod, we see ano he sha p downwa d end, ollowed by a sligh upwa d mo e. Thus, he middle egime is chosen objec i ely, since i desc ibes he pe iod o he G ea Recession, while he wo egimes be o e i , and he wo a e i , a e decided by subse s ha di e in e ms o p ice ola ili y. Viewed oge he , hese i e egimes a e dissimila in e ms o he economic condi ions d i ing he oil ma ke . Table 2. Da a se ime pe iods. Regime Begin End Numbe o Mon hs P e-Bubble Janua y 1986 Decembe 2001 180 Bubble Fo ming Janua y 2001 No embe 2007 83 C isis Decembe 2007 May 2009 18 A e C isis June 2009 May 2013 48 Recen June 2013 Decembe 2020 91 We used economic easoning o cha ac e ize he subdi isions p oposed. Ou economic easoning was d i en by a ca e ul and ealis ic assessmen o impo an de elopmen s in he global oil ma ke . The i e pe iods chosen and desc ibed as: P e-Bubble, Bubble Fo ming, C isis, A e C isis, and Recen , we e de e mined by s a ing wi h he well- de ined pe iod called he Global Financial C isis. The e is much consensus ha his pe iod s a ed in Decembe 2007 and ended in May 2009. Using his pe iod as an ancho , we chose a p e-c isis and a pos c isis pe iod, and u he analysis o hese wo b oad ca ego ies guided us owa ds subdi iding each o hese in o wo, because nume ous au ho s we ci e in he pape ha e desc ibed he eme gence o p e-c isis bubbles as well as pos -c isis de elopmen s in acking. This pape ’s claim is: by p oposing i e economic egimes and building i e boos ed and o e - i ed neu al ne wo ks o cap u e he exac ela ionships be ween spo oil p ices and oil da a ela ed o hose p ices, we will ecei e mo e insigh ul esul s han i we had jus analyzed a single pe iod. This analysis shows ha , while he inpu s in o an accu a e neu al ne wo k can emain he same, he impac o each a iable can change conside ably du ing di e en egimes, and he insigh s om ou speci ic egimes a e mo e use ul han he esul s om he whole sample pe iod. We do no claim ha we ha e p oposed he mos op imal i e egimes among all possible combina ions. Suppo o ou app oach is p o ided by s udies such as Balcila and Ozdemi ’s (2013) and Zhu e al.’s (2017). Balcila and Ozdemi (2013) in es iga ed he egime-swi ching causal nexus be ween U.S. oil and equi y ma ke s. The da a used we e he log e u ns o mon hly c ude oil u u es con ac s aded on he New Yo k Me can ile Exchange and a sub-g ouping o he S&P500 index. Thei sample pe iod co e ed he 1991 o 2011 pe iod, which includes se e al impo an de elopmen s in he oil sec o . These include he Gul C isis in 1991, he Asian inancial c ises in 1997–98, he 9/11 Te o is A ack in 2001, he J. Risk Financial Manag. 2021,14, 391 5 o 13 second Gul Wa in 2003, and inally, he global inancial c isis in 2007–9. The au ho s de i ed ou egimes and s udied ela ionships be ween he p ice o oil and equi y ma ke s. Zhu e al. (2017) conside ed he impac o oil supply and demand shocks on s ock e u ns du ing he pe iod o 1985 o 2015, using a wo- egime Ma ko egime-swi ching model. Unlike Balcila e al., Zhu e al. examined he asymme ic e ec s o oil supply and demand shocks on s ock e u ns in high- and low- ola ili y s a es. Thei wo- egime Ma ko egime-swi ching model lends suppo o he unde lying asymme ic esponses o s ock e u ns o oil supply and demand shocks o bo h oil-impo ing and oil-expo ing coun ies. Ou use o i e egimes is inspi ed by bo h Balcila e al. and Zhu e al., who o e empi ical e idence ha egimes di e , and hus ocusing on a numbe o egimes will en ich ou unde s anding. While hey applied a wo- egime Ma ko me hodology, we apply in his pape a neu al ne wo k me hodology, and ou da a a e ex ended o include oil da a du ing he pe iod o 2016 o 2020. Inc easing he numbe o egimes subs an ially beyond 5, say o 10 o 15, dilu es he impac o he majo e en s ha occu in ma ke s, and diminishes he abili y o each meaning ul conclusions. Simply pu , oo much de ail may obscu e c i ical ela ionships. Figu e 1shows his g aphically, along wi h he i e pe iods. In Figu es 2–5, g aphical ep esen a ions o each o he scaled inpu a iables a e depic ed. Ne impo s shows a g ea amoun o in e -week ola ili y and an unde lying in e ed pa abola shape, inc easing p io o he GFC, and dec easing a e wa ds. P oduc ion mo es in an upwa d pa abolic shape, wi h pe iodic seasonal d ops obse ed whe e oil demand slows. Following he GFC, p oduc ion inc eased signi ican ly, hen ook a dip. S ocks also exhibi a seasonal pa e n, wi h a sha p inc ease obse ed in he ecen ime pe iod. Consump ion shows a seasonal inc easing pa e n ha d opped du ing he GFC and has been g adually inc easing since hen. J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 5 o 13 a sub-g ouping o he S&P500 index. Thei sample pe iod co e ed he 1991 o 2011 pe iod, which includes se e al impo an de elopmen s in he oil sec o . These include he Gul C isis in 1991, he Asian inancial c ises in 1997–98, he 9/11 Te o is A ack in 2001, he second Gul Wa in 2003, and inally, he global inancial c isis in 2007–9. The au ho s de i ed ou egimes and s udied ela ionships be ween he p ice o oil and equi y ma - ke s. Zhu e al. (2017) conside ed he impac o oil supply and demand shocks on s ock e u ns du ing he pe iod o 1985 o 2015, using a wo- egime Ma ko egime-swi ching model. Unlike Balcila e al., Zhu e al. examined he asymme ic e ec s o oil supply and demand shocks on s ock e u ns in high- and low- ola ili y s a es. Thei wo- egime Ma - ko egime-swi ching model lends suppo o he unde lying asymme ic esponses o s ock e u ns o oil supply and demand shocks o bo h oil-impo ing and oil-expo ing coun ies. Ou use o i e egimes is inspi ed by bo h Balcila e al. and Zhu e al., who o e empi ical e idence ha egimes di e , and hus ocusing on a numbe o egimes will en ich ou unde s anding. While hey applied a wo- egime Ma ko me hodology, we apply in his pape a neu al ne wo k me hodology, and ou da a a e ex ended o in- clude oil da a du ing he pe iod o 2016 o 2020. Inc easing he numbe o egimes sub- s an ially beyond 5, say o 10 o 15, dilu es he impac o he majo e en s ha occu in ma ke s, and diminishes he abili y o each meaning ul conclusions. Simply pu , oo much de ail may obscu e c i ical ela ionships. Figu e 1 shows his g aphically, along wi h he i e pe iods. In Figu es 2–5, g aphical ep esen a ions o each o he scaled inpu a iables a e depic ed. Ne impo s shows a g ea amoun o in e -week ola ili y and an unde lying in e ed pa abola shape, inc eas- ing p io o he GFC, and dec easing a e wa ds. P oduc ion mo es in an upwa d pa a- bolic shape, wi h pe iodic seasonal d ops obse ed whe e oil demand slows. Following he GFC, p oduc ion inc eased signi ican ly, hen ook a dip. S ocks also exhibi a seasonal pa e n, wi h a sha p inc ease obse ed in he ecen ime pe iod. Consump ion shows a seasonal inc easing pa e n ha d opped du ing he GFC and has been g adually inc eas- ing since hen. Figu e 1. Spo p ices o e i e egimes. 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Recen Recen Scaled Spo P ice du ing Regimes Figu e 1. Spo p ices o e i e egimes. J. Risk Financial Manag. 2021,14, 391 6 o 13 J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 6 o 13 Figu e 2. Ne impo s. Figu e 3. P oduc ion. Figu e 4. S ocks. 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Recen Recen Scaled Ne Impo s du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled P oduc ion du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled S ocks Excluding SPR du ing Regimes Figu e 2. Ne impo s. J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 6 o 13 Figu e 2. Ne impo s. Figu e 3. P oduc ion. Figu e 4. S ocks. 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Recen Recen Scaled Ne Impo s du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled P oduc ion du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled S ocks Excluding SPR du ing Regimes Figu e 3. P oduc ion. J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 6 o 13 Figu e 2. Ne impo s. Figu e 3. P oduc ion. Figu e 4. S ocks. 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Recen Recen Scaled Ne Impo s du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled P oduc ion du ing Regimes 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled S ocks Excluding SPR du ing Regimes Figu e 4. S ocks. J. Risk Financial Manag. 2021,14, 391 7 o 13 J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 7 o 13 Figu e 5. Consump ion. In Table 3, co ela ions be ween he inpu a iables and he spo p ices du ing each o he ime pe iods a e shown. Bold alues highligh he inpu wi h he highes co ela ion in a gi en egime. As a i s indica ion o he shi ing ela ionships, we see ha he se o nega i e co ela ions is no consis en om pe iod o pe iod. No inpu is always posi i e o nega i e. P oduc ion is nega i ely co ela ed wi h spo p ices, excep in he pe iod ol- lowing he c isis. Consump ion has a posi i e co ela ion wi h he spo p ices be o e and h ough he c isis pe iod, hen u ns nega i e o bo h pe iods ollowing he c isis. The amoun o oil held in s ocks has a nega i e co ela ion wi h he spo p ices excep when he bubble was beginning o o m, while ne impo s ha e a posi i e ela ionship wi h he spo p ices excep du ing he pe iod immedia ely a e he c isis, whe e he ela ionship u ned sha ply nega i e. These shi ing co ela ion ela ionships a e an indica ion ha he impo ance o each o hese a iables in ela ion o spo p ices may change wi hin each o he neu al ne wo k models. Table 3. Co ela ions be ween inpu a iables and he spo p ice, pe pe iod. P oduc ion Consump ion S ocks Ne Impo s P e-Bubble −0.2910 0.2516 −0.4449 0.2274 BubbleFo ming −0.7807 0.4873 0.2760 0.6817 C isis −0.1076 0.4285 −0.7961 0.4046 A e C isis 0.4442 −0.2450 −0.4839 −0.5623 Recen −0.4035 0.0670 −0.8750 0.4553 3. Models Each o he da a se s was used o build a neu al ne wo k wi h IBM’s Modele 17.0 da a mining so wa e. Each model had i e inpu a iables and one a ge . The mul i- pe cep on ne wo ks had one hidden laye . Since he objec i e was o build models ha we e based as p ecisely as possible on each da a se , boos ing was used o enhance he model accu acy. Fo eplicabili y, he same seed o 229,176,228 was used o each o he ne wo ks. The models we e se o s op a e a maximum o 15 min, hough each ained o unde a minu e. In Modele , boos ing is done by c ea ing an ensemble model. Tha is, a sequence o en models is buil . In his g oup o componen models, each one is buil based on he en i e da a se o he gi en pe iod. Be o e building he nex successi e componen model, all he ows a e weigh ed based on he esiduals om he immedia ely p e ious compo- nen model. Rows wi h la ge esiduals ecei e ela i ely highe analysis weigh s in o de o o ce he ollowing componen model o o ecas hese pa icula eco ds accu a ely. 0 0.2 0.4 0.6 0.8 1 1.2 P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble P e-Bubble BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming BubbleFo ming C isis C isis A e C isis A e C isis A e C isis A e C isis Recen Recen Recen Recen Recen Recen Recen Scaled Consump ion du ing Regimes Figu e 5. Consump ion. In Table 3, co ela ions be ween he inpu a iables and he spo p ices du ing each o he ime pe iods a e shown. Bold alues highligh he inpu wi h he highes co ela ion in a gi en egime. As a i s indica ion o he shi ing ela ionships, we see ha he se o nega i e co ela ions is no consis en om pe iod o pe iod. No inpu is always posi i e o nega i e. P oduc ion is nega i ely co ela ed wi h spo p ices, excep in he pe iod ollowing he c isis. Consump ion has a posi i e co ela ion wi h he spo p ices be o e and h ough he c isis pe iod, hen u ns nega i e o bo h pe iods ollowing he c isis. The amoun o oil held in s ocks has a nega i e co ela ion wi h he spo p ices excep when he bubble was beginning o o m, while ne impo s ha e a posi i e ela ionship wi h he spo p ices excep du ing he pe iod immedia ely a e he c isis, whe e he ela ionship u ned sha ply nega i e. These shi ing co ela ion ela ionships a e an indica ion ha he impo ance o each o hese a iables in ela ion o spo p ices may change wi hin each o he neu al ne wo k models. Table 3. Co ela ions be ween inpu a iables and he spo p ice, pe pe iod. P oduc ion Consump ion S ocks Ne Impo s P e-Bubble −0.2910 0.2516 −0.4449 0.2274 BubbleFo ming −0.7807 0.4873 0.2760 0.6817 C isis −0.1076 0.4285 −0.7961 0.4046 A e C isis 0.4442 −0.2450 −0.4839 −0.5623 Recen −0.4035 0.0670 −0.8750 0.4553 3. Models Each o he da a se s was used o build a neu al ne wo k wi h IBM’s Modele 17.0 da a mining so wa e. Each model had i e inpu a iables and one a ge . The mul i- pe cep on ne wo ks had one hidden laye . Since he objec i e was o build models ha we e based as p ecisely as possible on each da a se , boos ing was used o enhance he model accu acy. Fo eplicabili y, he same seed o 229,176,228 was used o each o he ne wo ks. The models we e se o s op a e a maximum o 15 min, hough each ained o unde a minu e. In Modele , boos ing is done by c ea ing an ensemble model. Tha is, a sequence o en models is buil . In his g oup o componen models, each one is buil based on he en i e da a se o he gi en pe iod. Be o e building he nex successi e componen model, all he ows a e weigh ed based on he esiduals om he immedia ely p e ious componen model. Rows wi h la ge esiduals ecei e ela i ely highe analysis weigh s in o de o o ce he ollowing componen model o o ecas hese pa icula eco ds accu a ely. The componen models oge he o m he ensemble model. This ensemble model hen sco es J. Risk Financial Manag. 2021,14, 391 8 o 13 eco ds by using a combining ule. This combining ule assigns o he a ge he alue ha has he highes p obabili y mos o en ac oss he componen models. This is done o help he ne wo k o mos accu a ely mi o he ela ionships wi hin ha inancial egime o he da a. In Modele , all de elopmen s occu h ough nodes, and hey low om he i s node o he las . Nodes a e added o he s eam by d agging up in o he s eam he ype o node wished o om he possible se a he bo om o he s eam’s sc een a ea. Figu e 6 illus a es one o he s eams cons uc ed o his p oblem. I shows an Excel node connec ed o a Type node, hen o a Selec node ollowed by a neu al ne wo k node. The ne wo k gene a es a ained node (shown as a gold nugge ) h ough which i sends da a o gene a e a o ecas o each da ase ow. The las node, a able, is used o display he esul s. J. Risk Financial Manag. 2021, 14, x FOR PEER REVIEW 8 o 13 The componen models oge he o m he ensemble model. This ensemble model hen sco es eco ds by using a combining ule. This combining ule assigns o he a ge he alue ha has he highes p obabili y mos o en ac oss he componen models. This is done o help he ne wo k o mos accu a ely mi o he ela ionships wi hin ha inancial egime o he da a. In Modele , all de elopmen s occu h ough nodes, and hey low om he i s node o he las . Nodes a e added o he s eam by d agging up in o he s eam he ype o node wished o om he possible se a he bo om o he s eam’s sc een a ea. Figu e 6 illus- a es one o he s eams cons uc ed o his p oblem. I shows an Excel node connec ed o a Type node, hen o a Selec node ollowed by a neu al ne wo k node. The ne wo k gene a es a ained node (shown as a gold nugge ) h ough which i sends da a o gene a e a o ecas o each da ase ow. The las node, a able, is used o display he esul s. Figu e 6. Modele s eam. Se ings wi hin a node a e easily accessed by igh -clicking on he node and opening a submenu. The da a is i s ead in using an Excel node which o ms he connec ion o he speci ic da a se . I hen lows h ough a Type node whe e he ole each ield will play is speci ied. As Modele eads he da a, i also displays he la ges /smalles alues he a iables ake on, he ype o da a i is assigning o each a iable, and whe he o no he e a e any missing o un eadable da a in he se . The Type node, opened in i s Edi sc een, allows all hese se ing o be checked. F om he Type node, he da a a e ed h ough a Selec node, which allows he use o na ow he da a se o a speci ic ime pe iod. F om his node, he da a lows o he neu al ne wo k node. Wi hin he neu al ne wo k node, he se ings o model size and pu pose (boos ing) a e con olled. The e a e many possible se ings wi hin he Edi sc een o his node. In Objec i es, one can choose o build a s anda d model, o model o accu acy, o o model o s abili y. S opping ule limi s can be se o ime, o a numbe o cycles, o o accu acy. When using ei he boos ing o bagging, se ings unde Ensembles allow he use o speci y he ype o combining ule ha will be used o ei he con inuous o ca e- go y ype a ge s. Unde he Ad anced menu, he use can speci y he andom seed o he ne wo k un in o de o eplica e esul s, and whe he o no missing alues will be impu ed. The neu al ne wo k is hen execu ed and his gene a es a ained model. The ained model can be b owsed o iew he sensi i i y analysis and s uc u e o he model. Once a ained model is gene a ed, o he da a se s can be un h ough i o gene a e u u e o e- cas s o o compa e accu acies on a alida ion se . The ained model’s speci ic ou pu can be iewed by a aching a able o he ained node and unning he da a h ough i . A ne wo k simila o he one shown was de eloped o each o he i e da a se s. Figu e 6. Modele s eam. Se ings wi hin a node a e easily accessed by igh -clicking on he node and opening a submenu. The da a is i s ead in using an Excel node which o ms he connec ion o he speci ic da a se . I hen lows h ough a Type node whe e he ole each ield will play is speci ied. As Modele eads he da a, i also displays he la ges /smalles alues he a iables ake on, he ype o da a i is assigning o each a iable, and whe he o no he e a e any missing o un eadable da a in he se . The Type node, opened in i s Edi sc een, allows all hese se ing o be checked. F om he Type node, he da a a e ed h ough a Selec node, which allows he use o na ow he da a se o a speci ic ime pe iod. F om his node, he da a lows o he neu al ne wo k node. Wi hin he neu al ne wo k node, he se ings o model size and pu pose (boos ing) a e con olled. The e a e many possible se ings wi hin he Edi sc een o his node. In Objec i es, one can choose o build a s anda d model, o model o accu acy, o o model o s abili y. S opping ule limi s can be se o ime, o a numbe o cycles, o o accu acy. When using ei he boos ing o bagging, se ings unde Ensembles allow he use o speci y he ype o combining ule ha will be used o ei he con inuous o ca ego y ype a ge s. Unde he Ad anced menu, he use can speci y he andom seed o he ne wo k un in o de o eplica e esul s, and whe he o no missing alues will be impu ed. The neu al ne wo k is hen execu ed and his gene a es a ained model. The ained model can be b owsed o iew he sensi i i y analysis and s uc u e o he model. Once a ained model is gene a ed, o he da a se s can be un h ough i o gene a e u u e o ecas s o o compa e accu acies on a alida ion se . The ained model’s speci ic ou pu can be iewed by a aching a able o he ained node and unning he da a h ough i . A ne wo k simila o he one shown was de eloped o each o he i e da a se s.