Multi-country Analysis of Unemployment Rate Nowcasting During Covid-19 With Search Query Data

i Mas e Deg ee P og am in S a is ics and In o ma ion Managemen Mul i-coun y Analysis o Unemploymen Ra e Nowcas ing Du ing Co id-19 Wi h Sea ch Que y Da a A hu Hen ique Fe nandes Campos Disse a ion p esen ed as pa ial equi emen o ob aining he Mas e Deg ee P og am in S a is ics and In o ma ion Managemen NOVA In o ma ion Managemen School Ins i u o Supe io de Es a ís ica e Ges ão de In o mação Uni e sidade No a de Lisboa MEGI i NOVA In o ma ion Managemen School Ins i u o Supe io de Es a ís ica e Ges ão de In o mação Uni e sidade No a de Lisboa MULTI-COUNTRY ANALYSIS OF UNEMPLOYMENT RATE NOWCASTING DURING COVID-19 WITH SEARCH QUERY DATA By A hu Hen ique Fe nandes Campos Mas e Thesis p esen ed as pa ial equi emen o ob aining he Mas e ’s deg ee in S a is ics and In o ma ion Managemen , wi h a specializa ion in Risk Analysis and Managemen . Co-Supe iso : B uno Damásio Co-Supe iso : Flá io Pinhei o Feb ua y 2023 ii STATEMENT OF INTEGRITY I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no used plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he p ocess leading o i s elabo a ion. I u he decla e ha I ha e ully acknowledge he Rules o Conduc and Code o Hono om he NOVA In o ma ion Managemen School. São Paulo, 27 de e e ei o de 2023. iii ABSTRACT Nowcas ing me hods aim o p edic he p esen and he e y nea u u e and pas o ci cum en da a lag. As in e ne usage becomes ubiqui ous, mo e and mo e indi iduals use in e ne sea ch engines as decision-making ools; consequen ly, sea ch que y da a may be good p oxies o indi idual beha io , and hus a use ul nowcas ing p edic o a iable o many mac oeconomic indica o s. This s udy examines he po en ial o using Google T ends da a o nowcas unemploymen a e du ing he yea s o he Co id-19 pandemic ac oss six een coun ies by compa ing he pe o mance o ou al e na i e models wi h Google T ends da a agains a base au o eg essi e model, conside ing wo modelling aining windows, one limi ed o p e-Co id da a and he o he including 2020 da a. The esul s show ha sea ch que y da a lack obus ness and ha e a ying p edic i e powe , wi h he inclusion o 2020 da a in o he aining se p o iding a signi ican imp o emen o ou -o -sample o ecas ing accu acy. These indings indica e ha sea ch que y da a may ha e good p edic i e powe in some scena ios, bu may no be obus enough o eal-li e applica ions. KEYWORDS Economic o ecas ing; Nowcas ing; Unemploymen ; Google T ends. i INDEX 1. In oduc ion .................................................................................................................. 1 2. Li e a u e Re iew ......................................................................................................... 2 3. Da a............................................................................................................................... 5 4. Me hodology ................................................................................................................ 8 5. Resul s and Discussion .................................................................................................. 9 5.1. Scena io A: 2015–2019 .......................................................................................... 9 5.2. Scena io B: 2016–2020 ........................................................................................ 13 6. Conclusion .................................................................................................................. 19 Bibliog aphical Re e ences .............................................................................................. 20 Appendix A: Pas and Cu en Google T ends Sampling ................................................. 23 LIST OF FIGURES Figu e 5.1 – AIC and RMSE alues o 2020 and 2021 o he p e e ed models, Scena io A .. 13 Figu e 5.2 – AIC and RMSE alues o 2021 o he p e e ed models, Scena io B ................... 15 Figu e A.3 – Mean, 10 h and 90 h pe cen ile alues o A s & En e ainmen in he Uni ed S a es unde pas beha io , n=9 ...................................................................................... 23 Figu e A.4 – Mean, 10 h and 90 h pe cen ile alues o A s & En e ainmen in he Uni ed S a es unde cu en beha io be o e no icing any change, n=9 .................................... 24 Figu e A.5 – Mean, 10 h and 90 h pe cen ile alues o A s & En e ainmen in he Uni ed S a es unde cu en beha io a e p uning, n=79 ......................................................... 25 i LIST OF TABLES Table 3.1 – Da a sou ces and ype o ime se ies a ailable o unemploymen a e, pe coun y………………………………………………………………………………………………………………………….6 Table 5.1 – AIC alues pe model pe a iable pe coun y, Scena io A .................................... 9 Table 5.2 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2020, Scena io A ........................................................................................................................................ 10 Table 5.3 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2021, Scena io A ........................................................................................................................................ 11 Table 5.4 – AIC alues pe model pe a iable pe coun y, Scena io B .................................. 14 Table 5.5 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2021, Scena io B ........................................................................................................................................ 15 Table 5.6 – Ou -o -sample o ecas RMSE di e ences and pe cen age di e ences be ween Scena io B and Scena io A equi alen models o 2021. ................................................. 17 1 1. INTRODUCTION F om pe sonal budge ing and business planning o go e nmen policymaking, h oughou many ace s o mode n socie y, indi iduals and ins i u ions alike ely on o icial s a is ics o decision-making. Howe e , gi en he wo k in ol ed in da a collec ion, ea men , and analysis, all o icial eleases a e bound o be eleased wi h a delay (i.e., esul s o Feb ua y may only be a ailable in Ma ch). When accu a e p esen in o ma ion is impe a i e, i p o es necessa y o o ecas he p esen — he “now”. Nowcas ing aims o o ecas no he medium- o long- e m u u e, bu wha is happening igh now. O en, s a is icians include mo e ecen da a in o a amewo k; o ins ance, o p edic a mon hly indica o , i may be use ul o include esul s om a weekly su ey, o e en a co a ia e (i.e., a p edic o ) wi h he same elease equency bu di e en elease window. O e all, he mo o is: imely in o ma ion is s a egic. Wi h he widesp ead adop ion o he In e ne and he echnological ad ancemen s ha ollowed he digi al age, a as amoun o public da a became eadily a ailable online. One such da a sou ce, he sea ch que y agg ega o Google T ends, allows use s o ack he in e es in speci ic opics and ca ego ies ac oss geog aphy and ime. Unde he hypo hesis ha sea ch engine que ies a e a easonable p oxy o ac ual beha io s (e.g., pu chasing habi s), moni o ing changes in hese que ies migh gi e insigh in o cu en e en s and condi ions ha migh o he wise ake weeks o mon hs o be e lec ed in o icial s a is ics. Thus, should said hypo hesis p o e co ec , hen he po en ial exis s in using Google T ends (GT) da a o e ine nowcas ing esul s and make be e decisions as e . Since he seminal wo ks on he opic by Choi and Va ian (2009, 2012), many esea che s ha e explo ed he idea u he , o a ying deg ees o success. One cons an , howe e , has been he p e alence o s udies ocused on he de eloped wo ld, e en hough de eloping coun ies o en su e om longe elease lags o less eliable co a ia e da a al oge he . This s udy ocus on explo ing he po en ial o GT da a o nowcas ing unemploymen a e in 16 coun ies including high- and low-income economies. These a e Aus alia, B azil, Canada, Chile, Ge many, I aly, Japan, Mexico, he Ne he lands, Po ugal, Sou h Ko ea, Swi ze land, Tu key, he Uni ed Kingdom, he Uni ed S a es, and U uguay. These coun ies ha e easy- o-access mon hly unemploymen da a eleases om p ima y sou ces and, apa om U uguay, a e among he 50 la ges economies by nominal GDP. The disse a ion is o ganized as ollows. Chap e 2 con ains an o e iew o cu en li e a u e on he opics o nowcas ing and o ecas ing wi h sea ch que y da a. Chap e 3 desc ibes he da a and i s collec ion. Chap e 4 goes in o he nowcas ing me hodology. Chap e 5 p esen s and discusses he nowcas ing esul s pe scena io. Finally, Chap e 6 concludes he wo k wi h a b ie e lec ion upon he esul s and sugges ions o u he explo a ion o he heme. 2 2. LITERATURE REVIEW While back a i s launch Google T ends was pe cei ed as a ool o webmas e s and ma ke e s alike o sea ch engine op imiza ion pu poses, s a ing in 2008 Google Inc. and independen au ho s published pape s on applying sea ch engine agg ega ed da a o scien i ic esea ch in epidemiology (Polg een e al. 2008) and economic nowcas ing (Choi & Va ian, 2009). In Sep embe 2008, Google Inc. eleased Google Flu T ends (GFT), a lu nowcas ing se ice ueled by Google T ends que ies. Un il i s shu down on Augus 09, 2015, GFT was he opic o many discussions ega ding i s p edic i e powe and u ili y as an ou b eak p edic ion ool. No ably, Olson e al. (2013) ound ha , e en wi h he 2009 e ised me hodology, GFT ell sho by 52% in i s in luenza-like in ec ions p edic ion o New Yo k Ci y du ing he 2009 A/H1N1 pandemic. Bu , in a ecen u n o e en s, Kandula and Shaman (2019) look a new su eillance da a o ee alua e he GFT es ima ion e o s and c ea e a andom o es eg ession model wi h GFT a es ha see an e o educ ion o 80% o he 2012/13 season o iginal p edic ion, sugges ing a ee alua ion o sea ch que y usage as p edic o a iables in in luenza o ecas sys ems. In economics and business, howe e , ecep ion o GT da a-powe ed o ecas s and nowcas s has been mo e a o able, seeing use o ins ance in p edic ing ou ism in lows (A ola, Pin o & de Ped aza Ga cía, 2015) and demand (Sili e s o s & Wochne , 2018), suicide occu ences (K is ou ek, Moa & P eis, 2016), and ashion consume beha io o a big playe in he indus y (Sil a e al., 2019). In he opic o inancial ma ke s, GT da a ha e been used o p edic ing down u n s ock ma ke mo es (P eis, Moa & S anley, 2013), o eign exchange a es (Bulu , 2017), di ec ion o opening s ock p ices (Hu e al., 2018) and acc uable e u ns on p ecious me als (Salisu, Ogbonna & Adewuyi, 2020). Some au ho s a e less en husias ic abou he p ospec o Google T ends as a p edic o . Nagao, Takeda and Tanaka (2019) sugges ha GT da a-d i en models may lack obus ness and a e dependen on da a equency and seasonali y adjus men s wi h no consis ency ega ding whe he hey would imp o e o educe accu acy. Schae , Kou en zes and Fildes (2019) ind ha es ablished o ecas ing benchma ks ou pe o m hose wi h GT da a and social ne wo k in o ma ion in o ecas ing ideo game sales and co po a e online ideo iews, al hough he au ho s acknowledge hey a e limi ing he analysis o linea models. When i comes o a ocus on mac oeconomic a iables, he e seems o be a na u al endency o s udy dependen a iables closely ela ed o indi idual beha io , namely p i a e consump ion, and unemploymen ; Vosen and Schmid (2011), Choi and Va ian (2012), Vosen and Schimid (2012), Ca iè e-Swallow and Labbé (2013), and Woo and Owen (2019) look a he o me , while Choi and Va ian (2009), Ba ei a, Godinho and Melo (2013), Fondeu and Ka amé (2013), Vicen e, López- Menéndez and Pé ez (2015) and Nacca a o e al. (2018) a he la e . As GT da a pe ains mos ly o indi idual sea ch que ies, i is expec ed ha o ecas ing models o hose a iables would bene i om such da a. Fewe s udies, such as Ma cellino and Schumache (2010), Kuzin, Ma cellino and Schumache (2011), and Ban is, Clemen s and U quha (2021), look a GDP g ow h a e. As o loca ion, he economies s udied, o he mos pa , a e pa o he geopoli ical so-called de eloped wo ld; he excep ions being B azil (Ban is, Clemen s & U quha , 2021), Chile (Ca iè e- Swallow & Labbé, 2013) and Hong Kong, China (Choi & Va ian, 2012). Ca iè e-Swallow and Labbé (2013) poin ou ha good nowcas ing me hods a e e en mo e impo an in de eloping coun ies as 9 5. RESULTS AND DISCUSSION 5.1. SCENARIO A: 2015–2019 This scena io sepa a es p e-Co id and Co id-19 da a in o aining and es ing se s espec i ely. Should he al e na i e models be obus , hen hey should p oduce mo e accu a e 2020 and 2021 o ecas s han he base model. Fi e ou o 12 coun ies wi h unadjus ed ime se ies and six ou o 12 coun ies wi h seasonally adjus ed ime se ies had al e na i e models wi h lowe AIC han he base model, o a o al o se en coun ies: I aly, Japan, Po ugal, Sou h Ko ea, he Uni ed Kingdom, he Uni ed S a es, and U uguay. Table 5.1 p esen s he AIC alues o each model. Table 5.1 – AIC alues pe model pe a iable pe coun y, Scena io A Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U -196.56 -190.26 -191.17 -192.77 -193.96 SA -212.16 -208.35 -209.86 -208.76 -210.35 B azil U -223.78 -185.65 -181.83 -220.90 -217.50 Canada U -175.88 -172.53 -168.55 -172.79 -168.84 Chile U -178.74 -178.07 -175.41 -176.11 -173.48 SA -178.97 -178.34 -175.85 -176.43 -173.97 Ge many U -328.09 -322.74 -318.86 -325.59 -321.66 I aly SA -112.84 -99.29 -101.13 -114.48 -116.60 Japan U -211.99 -201.84 -199.11 -212.59 -209.27 SA -212.64 -209.97 -208.47 -216.23 -212.65 Mexico SA -168.22 -148.83 -144.94 -164.38 -160.62 Ne he lands U -232.18 -228.24 -226.55 -228.70 -226.78 SA -235.63 -233.99 -232.99 -233.53 -232.91 Po ugal U -175.49 -175.65 -172.03 -173.99 -170.35 SA -173.91 -176.00 -173.43 -174.58 -171.78 Sou h Ko ea U -145.08 -138.84 -137.06 -148.73 -145.23 SA -149.41 -147.39 -145.69 -153.10 -149.67 Swi ze land SA -211.66 -190.85 -187.97 -208.51 -206.92 Tu key SA -116.72 -97.09 -95.54 -113.22 -110.50 Uni ed Kingdom U -279.04 -276.83 -278.94 -275.13 -277.84 SA -280.08 -277.59 -280.09 -276.12 -279.33 Uni ed S a es U -210.75 -205.17 -207.09 -211.44 -211.54 SA -190.77 -191.32 -195.56 -192.48 -195.01 U uguay U -17.71 -13.36 -11.72 -22.68 -20.77 U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. Bold and unde sco ed alues indica e al e na i e models ha a e a p e e ed model. 10 A leas one al e na i e model bea s he base model in ou -o -sample pe o mance in eigh coun ies o 2020 da a and in nine coun ies o 2021 da a: Aus alia (2020, 2021), Canada (2020, 2021), Chile (2020, 2021), Ge many (2020, 2021), Japan (2020), Mexico (2020, 2021), he Ne he lands (2020, 2021), Tu key (2020, 2021), Sou h Ko ea (2021), and U uguay (2021). When limi ing he da a o p e e ed models, he e a e wo cases in which he p e e ed model ou pe o ms he base model in ou -o -sample pe o mance: he seasonally adjus ed se ies o Japan in 2020 and he unadjus ed se ies o U uguay in 2021, wi h espec i e changes in RMSE o -1,6% and -3,9%. On he o he end o he scale, he p e e ed model in he seasonally adjus ed se ies o I aly p oduces e y poo o ecas s compa ed o he base, wi h inc eases in RMSE o 193,2% in 2020 and 154,4% in 2021. Meanwhile, ano he al e na i e model wi h lowe AIC han he base a es be e han he p e e ed model, educing he di e ence in RMSE e sus he base model o 11,7% and 2,5% in 2020 and 2021 espec i ely. Excluding I aly as an ou lie , in 2020 da a he pe cen age di e ence in RMSE goes om -1,6% o 27,5%, while in 2021 da a i anges om -3,9% o 24,1%. Table 5.2 and Table 5.3 p esen he ou -o -sample o ecas RMSE alues o 2020 and 2021 espec i ely. Table 5.2 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2020, Scena io A Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U 0.46729 0.41738 0.44490 0.46807 0.50334 -10.68% -4.83% 0.17% 7.72% SA 0.46722 0.43739 0.46905 0.47175 0.50447 -6.38% 0.39% 0.97% 7.97% B azil U 0.30873 0.64435 0.70025 0.37633 0.38997 108.71% 126.82% 21.90% 26.31% Canada U 2.11369 1.91580 1.89169 2.12371 2.13443 -9.36% -10.50% 0.47% 0.98% Chile U 0.82059 0.81744 0.77748 0.83507 0.80104 -0.38% -5.25% 1.76% -2.38% SA 0.78695 0.77988 0.74368 0.80371 0.77472 -0.90% -5.50% 2.13% -1.55% Ge many U 0.31225 0.28212 0.28065 0.31491 0.31277 -9.65% -10.12% 0.85% 0.17% I aly SA 0.94008 1.82280 3.15852 1.05023 2.75601 93.90% 235.98% 11.72% 193.17% Japan U 0.14343 0.13911 0.14172 0.14585 0.13323 -3.01% -1.19% 1.68% -7.11% SA 0.16064 0.14468 0.16700 0.15801 0.16106 -9.94% 3.96% -1.64% 0.26% Mexico SA 0.52435 0.64092 0.63790 0.52890 0.52415 11 Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 22.23% 21.65% 0.87% -0.04% Ne he lands U 0.31453 0.278252 0.279894 0.31076 0.31044 -11.53% -11.01% -1.20% -1.30% SA 0.29893 0.27186 0.28046 0.28904 0.29979 -9.06% -6.18% -3.31% 0.29% Po ugal U 0.56934 0.60007 0.61227 0.59183 0.59781 5.40% 7.54% 3.95% 5.00% SA 0.55292 0.60245 0.63681 0.59573 0.62312 8.96% 15.17% 7.74% 12.69% Sou h Ko ea U 0.34768 0.49142 0.52016 0.43853 0.45705 41.34% 49.61% 26.13% 31.46% SA 0.28848 0.40153 0.42750 0.36783 0.38137 39.19% 48.19% 27.50% 32.20% Swi ze land SA 0.13944 0.21358 0.24823 0.18268 0.23603 53.17% 78.03% 31.01% 69.27% Tu key SA 0.76704 0.74723 0.88406 0.86881 0.97673 -2.58% 15.26% 13.27% 27.34% Uni ed Kingdom U 0.17210 0.18153 0.17881 0.17353 0.17261 5.48% 3.90% 0.83% 0.30% SA 0.17141 0.17992 0.17615 0.17162 0.17233 4.96% 2.77% 0.12% 0.54% Uni ed S a es U 3.44635 4.12511 3.96150 4.04447 4.07108 19.69% 14.95% 17.36% 18.13% SA 3.33147 4.19825 4.07052 4.04691 4.05539 26.02% 22.18% 21.48% 21.73% U uguay U 0.74317 0.92234 1.01664 0.77248 0.82870 24.11% 36.80% 3.95% 11.51% U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. Pe cen age di e ence alues a e in ela ion o he base model. Bold and unde sco ed alues indica e al e na i e models ha a e a p e e ed model. Table 5.3 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2021, Scena io A Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U 0.70517 0.60934 0.66429 0.71141 0.77040 -13.59% -5.80% 0.88% 9.25% SA 0.70060 0.62743 0.68573 0.69659 0.75177 -10.44% -2.12% -0.57% 7.30% B azil U 0.32060 0.84588 0.90028 0.44096 0.45682 163.84% 180.81% 37.54% 42.49% Canada U 2.01147 1.79371 1.76694 2.02401 2.03338 -10.83% -12.16% 0.62% 1.09% Chile U 0.99230 0.97980 0.97706 1.00165 1.00449 12 Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 -1.26% -1.54% 0.94% 1.23% SA 0.96462 0.94404 0.93942 0.97539 0.97671 -2.13% -2.61% 1.12% 1.25% Ge many U 0.35236 0.31285 0.31122 0.35408 0.35163 -11.21% -11.68% 0.49% -0.21% I aly SA 1.04137 1.70160 2.99813 1.06516 2.64925 63.40% 187.90% 2.28% 154.40% Japan U 0.19812 0.21534 0.22989 0.20430 0.21398 8.69% 16.04% 3.12% 8.01% SA 0.17263 0.18672 0.19648 0.17883 0.18322 8.16% 13.82% 3.59% 6.13% Mexico SA 0.56578 0.54537 0.54495 0.56812 0.57234 -3.61% -3.68% 0.41% 1.16% Ne he lands U 0.37383 0.36204 0.32729 0.39018 0.36053 -3.15% -12.45% 4.37% -3.56% SA 0.37494 0.41206 0.39819 0.42284 0.41850 9.90% 6.20% 12.78% 11.62% Po ugal U 0.56963 0.61858 0.63844 0.60836 0.61886 8.59% 12.08% 6.80% 8.64% SA 0.59537 0.66631 0.71772 0.65725 0.69840 11.92% 20.55% 10.39% 17.31% Sou h Ko ea U 0.52175 0.47096 0.47508 0.52715 0.53720 -9.74% -8.95% 1.03% 2.96% SA 0.44550 0.41665 0.43138 0.44863 0.45105 -6.48% -3.17% 0.70% 1.24% Swi ze land SA 0.15071 0.22922 0.26545 0.18945 0.23728 52.09% 76.13% 25.70% 57.44% Tu key SA 1.04308 0.90335 1.17248 1.05918 1.28279 -13.40% 12.41% 1.54% 22.98% Uni ed Kingdom U 0.25535 0.27667 0.28343 0.25747 0.26028 8.35% 10.99% 0.83% 1.93% SA 0.24850 0.27240 0.27934 0.24881 0.25431 9.62% 12.41% 0.13% 2.34% Uni ed S a es U 3.41673 4.11576 3.95774 4.03864 4.07116 20.46% 15.83% 18.20% 19.15% SA 3.26239 4.16751 4.04854 4.01225 4.02966 27.74% 24.10% 22.98% 23.52% U uguay U 1.06546 0.94425 0.98781 1.02420 1.11378 -11.38% -7.29% -3.87% 4.54% U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. Pe cen age di e ence alues a e in ela ion o he base model. Bold and unde sco ed alues indica e al e na i e models ha a e a p e e ed model. 13 Compa ing he ou -o -sample pe o mance o 2021 agains 2020, Canada, I aly, Mexico, Sou h Ko ea, he Uni ed S a es, and U uguay see a model ge mo e accu a e. In pa icula , he coun ies ha ha e hei p e e ed model imp o e in 2021 e sus 2020 a e Canada, I aly, and he Uni ed S a es, he las being he only o he h ee wi h a p e e ed al e na i e model. Figu e 5.1 plo s he AIC o he p e e ed models and hei espec i e RMSEs o 2020 and 2021 o ecas s. Figu e 5.1 – AIC and RMSE alues o 2020 and 2021 o he p e e ed models, Scena io A Wi h mo e han hal o he coun ies no ha ing sea ch que y da a as pa o hei p e e ed model, i sugges s ha sea ch que y models do no impa much in o ma ion ha is no al eady con ained in a simple au o eg essi e componen . To make ma e s wo se, only in wo cases he p e e ed al e na i e models ha e be e ou -o -sample pe o mance han he base model. E en hough hal he coun ies ha e a leas one al e na i e model ou pe o m he base in 2020 o 2021, sea ch que y da a may no be a eliable solu ion o imp o e nowcas ing accu acy. This can be u he seen in he gene al dec ease in o ecas ing powe in 2021 e sus 2020, which also pu s in o ques ion he obus ness o he da a as p edic o a iables. 5.2. SCENARIO B: 2016–2020 Gi en ha in his scena io he aining window includes he spikes in unemploymen o en seen a he beginning o he Co id-19 pandemic, i is easonable o expec a la ge numbe o coun ies wi h al e na i e models as bes pe o ming han in Scena io A, as a pu ely au o eg essi e model is poo a o ecas ing ou lie s. Aus alia B azil Mexico Po ugal Tu key Uni ed S a es Japan (U) Japan (SA) Uni ed Kingdom I aly Canada Sou h Ko ea Ne he lands Swi ze land Ge many U uguay Chile 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 -350 -300 -250 -200 -150 -100 -50 0 RMSE AIC Jan/20–Dec/20 Jan/21–Dec/21 14 Indeed, eigh ou o 12 coun ies wi h unadjus ed ime se ies and se en ou o 12 coun ies wi h seasonally adjus ed ime se ies had al e na i e models ou pe o m he base model wi hou GT da a, o a o al o 10 coun ies: Canada, Chile, Ge many, I aly, Japan, he Ne he lands, Po ugal, he Uni ed Kingdom, he Uni ed S a es, and U uguay. In hose coun ies, models wi h AR(1) componen s and GT da a, wi h o wi hou lagged ca ego ies, had he lowes AIC alues in 11 ou o 15 cases, and lowe AIC alues han he base model in 14 ou o 15 cases. Table 5.4 p esen s he AIC alues o each model. Table 5.4 – AIC alues pe model pe a iable pe coun y, Scena io B Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U -167.68 -165.46 -164.57 -164.44 -165.00 SA -166.97 -165.06 -163.65 -163.72 -163.35 B azil U -191.61 -149.59 -147.75 -189.93 -187.76 Canada U -22.41 -14.69 -78.14 -37.52 -123.71 Chile U -131.43 -102.15 -110.42 -127.57 -145.26 SA -134.47 -107.33 -115.62 -130.63 -148.69 Ge many U -234.21 -229.96 -242.08 -230.36 -246.62 I aly SA -83.27 -111.38 -109.39 -120.22 -116.28 Japan U -215.15 -206.85 -206.62 -216.42 -215.44 SA -215.13 -213.60 -212.84 -217.77 -215.93 Mexico SA -138.48 -117.65 -115.64 -135.83 -134.35 Ne he lands U -205.40 -204.18 -200.86 -202.72 -199.53 SA -205.23 -206.45 -203.08 -204.50 -201.14 Po ugal U -133.68 -131.17 -137.72 -130.49 -136.65 SA -134.93 -132.39 -138.23 -131.34 -136.83 Sou h Ko ea U -131.52 -118.98 -115.15 -128.56 -124.62 SA -140.35 -132.19 -128.60 -138.33 -134.33 Swi ze land SA -209.13 -186.06 -182.86 -205.50 -203.30 Tu key SA -82.29 -72.42 -71.34 -80.56 -78.38 Uni ed Kingdom U -261.38 -257.39 -261.36 -257.65 -261.90 SA -261.25 -259.61 -263.16 -258.27 -261.97 Uni ed S a es U 49.93 -25.16 -43.81 -30.45 -65.32 SA 50.19 -24.25 -43.04 -29.06 -64.06 U uguay U -15.45 -7.74 -4.88 -19.09 -15.25 U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. Bold and unde sco ed alues indica e al e na i e models ha a e a p e e ed model. Ou -o -sample o ecas s o 2021 a e also be e han in Scena io A. The base model has he lowes RMSE in only h ee coun ies: B azil, Japan, and he Ne he lands, he las being one o he coun ies wi h an al e na i e model su passing he base in AIC. O he o he hi een coun ies, only he Uni ed Kingdom has jus one al e na i e model wi h lowe e o han he base; meanwhile, all al e na i e 15 models ou pe o m he base model o Sou h Ko ea, he Uni ed S a es, and U uguay. The bes imp o emen is seen in he seasonally adjus ed ime se ies o he Uni ed S a es wi h a d op o 62,3% in RMSE compa ed o he base model when using GT da a wi h lagged ca ego ies and an AR(1) componen . When looking only a he p e e ed models pe a iable pe coun y, 10 ou o 24 ou pe o m he base models while 5 unde pe o m agains hem, wi h he pe cen age di e ence in RMSE anging om - 62,3% o -4,0% and +1,5% o +29,3% espec i ely. Figu e 5.2 plo s he AIC o he p e e ed models and hei espec i e RMSEs o 2020 and 2021 o ecas s, and Table 5.5 p esen s all ou -o -sample o ecas esul s o 2021. Figu e 5.2 – AIC and RMSE alues o 2021 o he p e e ed models, Scena io B Table 5.5 – Ou -o -sample o ecas RMSE alues and pe cen age di e ences o 2021, Scena io B Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U 0.57642 0.60310 0.58848 0.56031 0.52376 29.06% 25.93% 19.91% 12.09% SA 0.59580 0.61586 0.60052 0.58119 0.54731 31.81% 28.53% 24.39% 17.14% B azil U 0.31901 0.65181 0.64579 0.32765 0.32206 111.13% 109.18% 6.13% 4.32% Canada U 1.53422 1.62754 0.99586 1.21164 0.73224 -23.00% -52.89% -42.68% -65.36% Chile U 0.64865 1.02168 1.04790 0.64939 0.57271 Aus alia B azil Mexico Po ugal Tu key Uni ed S a es Japan Uni ed Kingdom I aly Canada Sou h Ko ea (U) Sou h Ko ea (SA) Ne he lands (U) Ne he lands (SA) Swi ze land Ge many U uguay Chile 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -300 -250 -200 -150 -100 -50 0 RMSE AIC 16 Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 24.50% 27.70% -20.86% -30.21% SA 0.62036 0.99149 1.01037 0.61401 0.54275 25.99% 28.39% -21.98% -31.03% Ge many U 0.29249 0.31276 0.23685 0.29117 0.17960 0.16% -24.14% -6.75% -42.48% I aly SA 0.87485 0.38721 0.42414 0.49770 0.49429 -58.81% -54.88% -47.06% -47.42% Japan U 0.20656 0.21932 0.24769 0.21183 0.23429 52.91% 72.69% 47.69% 63.35% SA 0.18359 0.19177 0.21013 0.18630 0.20218 19.38% 30.81% 15.98% 25.86% Mexico SA 0.57279 0.55165 0.57220 0.56889 0.58640 5.21% 9.13% 8.49% 11.83% Ne he lands U 0.33971 0.39808 0.36674 0.38872 0.35175 26.56% 16.60% 23.59% 11.83% SA 0.35287 0.45624 0.43432 0.45518 0.43263 52.62% 45.29% 52.27% 44.73% Po ugal U 0.55916 0.55789 0.42579 0.53686 0.41498 -2.01% -25.21% -5.71% -27.11% SA 0.58790 0.58865 0.46397 0.57013 0.45418 6.46% -16.09% 3.11% -17.86% Sou h Ko ea U 0.52426 0.45848 0.45894 0.52203 0.52000 31.87% 32.00% 50.15% 49.56% SA 0.45577 0.42055 0.42900 0.45457 0.45475 45.78% 48.71% 57.57% 57.64% Swi ze land SA 0.15228 0.23391 0.23125 0.14400 0.13403 67.76% 65.84% 3.27% -3.88% Tu key SA 0.98884 0.94397 0.92488 1.05171 1.01818 23.07% 20.58% 37.11% 32.74% Uni ed Kingdom U 0.24321 0.28529 0.30061 0.24122 0.26242 65.77% 74.67% 40.16% 52.48% SA 0.26007 0.28228 0.29711 0.25876 0.27594 64.68% 73.33% 50.96% 60.98% Uni ed S a es U 3.17296 1.64167 1.39472 1.54176 1.22023 -52.37% -59.53% -55.26% -64.59% SA 3.14488 1.60175 1.37722 1.52074 1.18515 -51.92% -58.66% -54.35% -64.43% U uguay U 1.09094 0.96543 0.94779 1.04696 1.06386 29.91% 27.53% 40.88% 43.15% U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. Pe cen age di e ence alues a e in ela ion o he base model. Bold and unde sco ed alues indica e al e na i e models ha a e a p e e ed model. 17 Shi ing he aining window as o include 2020 da a has a e y posi i e impac on he o e all pe o mance o al e na i e models, be i du ing aining— om se en o 10 coun ies wi h p e e ed al e na i e models—o ou -o -sample o ecas ing in 2021— om nine o 13 coun ies wi h al e na i e models ou pe o ming he base. A signi ican imp o emen is seen when ocusing on he pe o mance o p e e ed models in 2021: only one model has lowe RMSE han he base in Scena io A, while in Scena io B ha numbe jumps o en. This imp o emen can also be seen in he pe cen age di e ence in RMSE as he lowes dec ease in e o in Scena io B (-4,0%) is al eady highe han he single dec ease in Scena io A (-3,9%). Fu he mo e, when compa ing he ou -o -sample o ecas pe o mance o 2021 be ween models om Scena io A and Scena io B, a leas one al e na i e model om he la e has lowe RMSE han i s equi alen in he o me in 19 ou o 24 se ies; in 11 o hose, all Scena io B al e na i e models ou pe o m hei Scena io A equi alen s. The only coun ies wi h Scena io A models ha bea hei Scena io B equi alen s a e Japan, he Uni ed Kingdom, and U uguay. The pe cen age di e ence in RMSE anges om -85,85% o +12,05%. Table 5.6 p esen s he RMSE alue di e ence and pe cen age di e ence be ween Scena io A and Scena io B equi alen models o 2021. Table 5.6 – Ou -o -sample o ecas RMSE di e ences and pe cen age di e ences be ween Scena io B and Scena io A equi alen models o 2021. Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 Aus alia U -0.12876 -0.00625 -0.07581 -0.15109 -0.24664 -18.26% -1.03% -11.41% -21.24% -32.01% SA -0.10481 -0.01157 -0.08520 -0.11540 -0.20445 -14.96% -1.84% -12.43% -16.57% -27.20% B azil U -0.00159 -0.19407 -0.25448 -0.11332 -0.13476 -0.50% -22.94% -28.27% -25.70% -29.50% Canada U -0.47725 -0.16617 -0.77108 -0.81237 -1.30114 -23.73% -9.26% -43.64% -40.14% -63.99% Chile U -0.34365 0.04188 0.07084 -0.35226 -0.43177 -34.63% 4.27% 7.25% -35.17% -42.98% SA -0.34426 0.04745 0.07095 -0.36138 -0.43396 -35.69% 5.03% 7.55% -37.05% -44.43% Ge many U -0.05987 -0.00009 -0.07436 -0.06292 -0.17203 -16.99% -0.03% -23.89% -17.77% -48.92% I aly SA -0.16652 -1.31439 -2.57398 -0.56746 -2.15496 -15.99% -77.24% -85.85% -53.27% -81.34% Japan U 0.00844 0.00398 0.01779 0.00754 0.02031 4.26% 1.85% 7.74% 3.69% 9.49% SA 0.01097 0.00505 0.01366 0.00748 0.01896 6.35% 2.70% 6.95% 4.18% 10.35% Mexico SA 0.00702 0.00629 0.02726 0.00077 0.01406 1.24% 1.15% 5.00% 0.13% 2.46% Ne he lands U -0.03412 0.03605 0.03944 -0.00146 -0.00878 18 Coun y Va iable Model 1 Model 2 Model 3 Model 4 Model 5 -9.13% 9.96% 12.05% -0.37% -2.43% SA -0.02207 0.04418 0.03613 0.03233 0.01414 -5.89% 10.72% 9.07% 7.65% 3.38% Po ugal U -0.01047 -0.06069 -0.21265 -0.07151 -0.20388 -1.84% -9.81% -33.31% -11.75% -32.94% SA -0.00747 -0.07767 -0.25374 -0.08711 -0.24422 -1.25% -11.66% -35.35% -13.25% -34.97% Sou h Ko ea U 0.00250 -0.01248 -0.01614 -0.00511 -0.01721 0.48% -2.65% -3.40% -0.97% -3.20% SA 0.01027 0.00390 -0.00238 0.00593 0.00370 2.30% 0.94% -0.55% 1.32% 0.82% Swi ze land SA 0.00157 0.00469 -0.03421 -0.04545 -0.10326 1.04% 2.05% -12.89% -23.99% -43.52% Tu key SA -0.05424 0.04061 -0.2476 -0.00746 -0.26461 -5.20% 4.50% -21.12% -0.70% -20.63% Uni ed Kingdom U -0.01214 0.00862 0.01718 -0.01625 0.00214 -4.75% 3.12% 6.06% -6.31% 0.82% SA 0.01157 0.00988 0.01778 0.00995 0.02163 4.66% 3.63% 6.36% 4.00% 8.50% Uni ed S a es U -0.24378 -2.47409 -2.56301 -2.49688 -2.85093 -7.13% -60.11% -64.76% -61.82% -70.03% SA -0.11751 -2.56577 -2.67132 -2.49151 -2.84451 -3.60% -61.57% -65.98% -62.10% -70.59% U uguay U 0.02548 0.02118 -0.04002 0.02276 -0.04992 2.39% 2.24% -4.05% 2.22% -4.48% U and SA indica e unadjus ed se ies and seasonally adjus ed se ies espec i ely. Model numbe ing ollows he numbe ing om he me hodology sec ion. All alues a e in ela ion o he Scena io A equi alen model. While hese esul s a e p omising, hey come a he expense o inco po a ing s ess da a in o he aining. I is s ill unclea wha d i es he imp o emen s seen; jus as hey may come om he models being ained on he same s ess scena io hey a e o ecas ing, i may also be he case ha sea ch que y da a is g owing in o ecas ing powe as mo e indi iduals sea ch o and apply o jobs h ough he In e ne . 25 Figu e A.5 – Mean, 10 h and 90 h pe cen ile alues o A s & En e ainmen in he Uni ed S a es unde cu en beha io a e p uning, n=79 40 50 60 70 80 90 100 In e es o e ime Da e 10 h o 90 h pe cen ile ange Mean