ucela i-LLM ukuthi ilingise imindeni yaseMelika engu-6,000 ephendula imibuzo mayelana nokwehla kwamandla emali? Amaphepha akamuva athola ukuthi amamodeli ezilimi amakhulu angakwazi ukuphindaphinda izimpendulo ezimaphakathi zenhlolovo enkulu yasendlini ibe phakathi kwephuzu lephesenti (Zarifhonarvar, 2026). Ngo-2020, i-Survey of Consumer Expectations (SCE) yabika izinga lokwehla kwamandla emali eliphakathi nonyaka owodwa ngaphambili elicishe libe ngu-3%. I-median ekhiqizwe i-LLM egqugquzelwe enabantu abangokoqobo kanye nomyalelo wokunqamula ulwazi: futhi cishe u-3%. Vala ngokwanele ukuthi ama-LLM abekwe njengenani eliphansi, elihambisana nemvamisa ephezulu kuhlolo lwe-SCE, Michigan, kanye ne-Survey of Professional Forecasters.
Ephepheni lakamuva, Ingabe Ama-LLM Angakwazi Ukulingisa Izinhlolovo Zasekhaya?ebhalwe ngokubambisana no-Ami Dalloul wase-University of Duisburg-Essen, sibheka umzuzu wesibili, ingxenye yokusabalalisa okungenzeka okukutshela ukuthi imodeli imele umbono owodwa noma inkulungwane. Kulapha lapho kunyamalala khona impumelelo ebonakalayo yocwaningo olusekelwe ku-LLM. Imodeli efanayo ye-Llama-3 eshaya imidiyeni ye-SCE ibe phakathi kwephuzu lephesenti ibeka u-95% wabaphenduli bayo abalingiswayo ngaphakathi kwewindi elingamaphesenti amabili. Izimpendulo zangempela ze-SCE zika-2020 zisukela cishe ku-25 kuye ku-27%. Ngamafuphi, isilinganiso silungile, kodwa abantu abasilandelayo abekho. Ngakho-ke ukusebenzisa ukulingisa ngezinkulungwane ezimbalwa zabantu be-LLM kubilisa kummeleli oyedwa.
Umfanekiso 1: Ukusatshalaliswa Kwenani Labantu Lomhlaba Wangempela kanye Nenhlolovo Yokwenziwa
Qaphela: Iphaneli yesokunxele ihlela ukuhlakazwa kwabaphenduli be-SCE abangabodwana bango-2020 ngokwesilinganiso sabo. Ukusabalalisa imisebe kubonisa izinkolelo ezihlukahlukene kubo bonke abaphendulayo. Iphaneli emaphakathi isebenzisa ukwakhiwa okufanayo ezimpendulweni zokwenziwa ezivela kumodeli ye-Llama-3.1-8B-Instruct eyalwe ngabantu abafana nokusatshalaliswa kwabantu kwe-SCE. I-scatter iwa ifike endaweni eseduze. Imodeli ibuyisela i-mean futhi ilahle konke okunye. Iphaneli engakwesokudla isebenzisa imodeli efanayo ye-Llama engafundiwe ngokunyuka kwe-gradient (GA). Imodeli engafundiwe ifinyelela ukuhlakazeka okungokoqobo futhi ayigoqi kumodi.
Ukugoqa kwemodi
Silinganise ama-LLM amahlanu (Llama-3-8B, Llama-3-70B, Claude-3.7-Sonnet, DeepSeek-V3, GPT-4o) ngokumelene ne-SCE, i-Michigan Survey, kanye ne-Survey of Professional Forecasters. Ocwaningweni lwabantu, ama-44 kuye kwangama-70% abaphendulile anikeza izimpendulo ezingaphezu kwamaphesenti ama-3 kude nempendulo yemodeli; kumasampula e-LLM, lelo sabelo empeleni linguziro.
Amakhambi ajwayelekile asuka ezincwadini zokulingisa inhlolovo awayithuthukisi le nkinga. Abantu abasuselwe ekubaleni abantu abanezici eziyinkimbinkimbi nezihlukahlukayo, imiyalelo yokunqanyulwa kolwazi (“awuzazi izehlakalo ngemva kukaJuni 2018”), kanye nokubeka ingcaca kokuthi “ungabheki izibalo” kushukumisela ukuthi konke kube ngokuzenzakalela ekusabalaliseni okufanayo okuncane. Imbangela okungenzeka ukuthi ama-LLM abona amathebula e-CPI, ukusakazwa kwezindaba kokukhishwa kwenhlolovo ye-FRBNY, kanye nezimpendulo zezemfundo enhlanganweni yabo yokuqeqesha. Icelwa ngokulindeleka kokwehla kwamandla emali okumaphakathi kwango-2020, imodeli ibuyisa idatha ebanjwe ngekhanda. Isisindo saleyo datha yokuqeqeshwa singaphezu kwanoma yini imiyalelo yokwaziswa eyicela ukuba iyenze.
Ukuvula ama-LLM
Uma izibalo ezishiwo ngekhanda ziyinkinga, okungase kube ukulungiswa ukuzisusa ezisindweni kunokucela imodeli ukuthi ibheke eceleni. Sisebenzise izindlela ezimbili zokungafundi ku-Llama-3.1-8B-Instruct, imodeli yomthombo ovulekile esivumela ukuthi siguqule izisindo zayo:
- I-Gradient Ascent (GA) ikhulisa ukulahleka kokuqagela kusethi yokukhohlwa yochungechunge lwe-CPI nezihlanganisi zenhlolovo, ngokulahleka okukhumbulayo ekucabangeni kwenhlolovo encane ukuze amandla ajwayelekile asinde.
- I-Negative Preference Optimization (NPO) iphatha ukukhohlwa isethi njengokuqedwa okungathandwayo futhi inciphise ukulahlekelwa okuthandwayo okunomkhawulo ngokumelene nemodeli yesithenjwa.
Idatha esiyicela imodeli ukuthi iyikhohlwe irekhodi lokwehla kwamandla emali ngokwalo: uchungechunge lwanyanga zonke lwe-CPI kanye nokushicilelwe okulindelwe ukwehla kwamandla emali okuvela kuhlolo lwe-FRBNY SCE kanye noMichigan. Umthelela ongafundi ekusabalaliseni impendulo ukuThebula 1.
Ithebula 1 Ukunemba Komsila Ngamasu Ahlukene Wokungafundi
Qaphela: Amasu angafundi okunciphisa ukuwa kwemodi. I-Gradient ascent (GA) iyindlela eqondiwe yokungafundi lapho imodeli icushwe kahle ukuze kwandiswe ukulahlekelwa kudathasethi yezibalo ze-CPI kuyilapho kuncishiswa ukulahlekelwa, noma kugcinwa (RT), kudathasethi yedatha ye-micro-survey. I-Negative preference optimization (NPO) iphatha izibalo ezisemthethweni njengamasampuli angalungile ukuze ajezise isizukulwane sawo kuyilapho kuphatha amasampula agciniwe (RT) njengephozithivu. Izimpendulo zenhlolovo yokwenziwa kokulindelwe yi-inflation njengamaphesenti ukusuka kumodi kanye nencazelo (kubakaki) ngaphakathi kwemigqomo yokufanisa ncamashi, ± 1, kanye>> 3 % ukuchezuka. Umsila Acc. ikala ukusondelana nebhentshimakhi yokuhlakazeka komsila we-FRBNY (> ± 3.0 = 44.38).
I-Llama-3 eyisisekelo (ehlanganisa ukungafundi okusekelwe ngokushesha) ikhiqiza imodi efana ncamashi nezimpendulo ezingu-92% futhi uziro uphendula ngaphezu kuka-3pp. Ukunemba komsila uma kuqhathaniswa nebhentshimakhi ye-SCE yama-44% ngakho-ke inguziro. Ngemuva kwe-GA, ukufana okuqondile kwehlela ku-24%, futhi u-43% wezimpendulo udlulela ku-±3pp; ukunemba komsila kufinyelela ku-97%. I-NPO iqhathaniswa no-37% no-43%, nokunemba komsila okungama-98%. Ngamanye amazwi, zombili izindlela zokungafundi zibonakala zibuyisela ukusabalalisa okungokoqobo.
Umfanekiso 2 Ukuhlakazwa kwama-LLM vs. Amamodeli Angafundi
Qaphela: Uhlangothi lwesokunxele luceba izilinganiso ze-kernel zokuminyana kwe-kernel okulindelwe ku-inflation ka-2020 kusukela ku-FRBNY SCE kanye nokuhluka okubili kwe-Llama-3 okuqeqeshwe ngezindlela zokungafundi, ukukhuphuka kwe-gradient (GA) kanye nokwenza kahle kokuthandwa okungekuhle (NPO). Kokubili okuhlukile okungafundiwe kumboza ububanzi lapho i-FRBNY SCE ibeka khona isisindo samathuba, nakuba zisalokhu zigxile kakhulu kunokulinganisa komuntu futhi zitshekele kancane ezindleleni eziphakeme. Uhlangothi olungakwesokudla luqhathanisa ama-KDE okulindelekile okukhiqizwa yi-LLM (GPT-4o, Llama-3, njll.) ne-FRBNY SCE ngo-2020. Amajika e-LLM (i-eksisi yesokunxele) ahlanganiswe ngokuqinile endaweni ewumngcingo, kuyilapho ijika le-FRBNY SCE lihlala libanzi kakhulu. Ama-LLM angakwazi ukufanisa ukuthambekela okumaphakathi nokho ahluleke ukukhiqiza kabusha ukusabalala kwezingxenye ezihlukene zedatha encane yocwaningo. Umkhawulokudonsa = 0.5 wawo wonke ama-KDE.
Ukuminyana kwe-kernel (Umfanekiso 2) kubonisa ukuthi amamodeli angekho eshalofini anqwabelanisa ubuningi bamathuba abe yi-spike encane eduze nendawo. Izinhlobonhlobo ezingafundiwe zisakaza insada phakathi kobubanzi lapho abantu abaphendula i-SCE bayibeka khona.
Ukulingisa isilingo esilawulwa ngokungahleliwe
Ukusabalalisa okubanzi kuyadingeka kodwa akwanele kuhlelo lokusebenza olugqugquzele iphepha lethu: ukuphindaphinda ama-RCT ohlolo ngezinguqulo zokwenziwa. Ama-RCT ayabiza. Ngemuva kokuphela kokuqoqwa kwedatha, umcwaningi akakwazi ukubuyela emuva ukuze ahlole ithiyori eyavela kamuva noma aguqule ukwelashwa. Ama-synthetic agents angasivumela ukuthi senze lokho kanye, uma ukuziphatha kwabo kufana nalokho okukhiqizwa abaphendulayo bangempela.
Ukuhlola lokhu, siphindaphinda i-RCT yomhlaba wangempela ka-Coibion, Gorodnichenko, and Weber (2022). Abaphenduli banikezwa ngokungahleliwe kwelinye lamaqembu amaningana: iqembu elilawulayo aliboni ulwazi, amaqembu amaningana okwelapha ngalinye lithola ucezu lwezomnotho oluhlukile (izinga langempela lokukhuphuka kwamandla emali, izinga le-Fed lika-2%, njll.), futhi iqembu le-placebo liboniswa okuqukethwe okungahlobene nokwehla kwamandla emali. Bonke abaphendulile baqale babike okulindelekile ngaphambi kokwehla kwamandla emali, bese bebona noma yini eyabelwe iqembu labo, bese bebika okulindelekile kwangemuva okusha. Umehluko phakathi kwengemuva nengaphambili isibuyekezo somphenduli.
Ukwelashwa kusebenza uma izibuyekezo zakhona zihluka ngokusobala kusukela kweqembu elilawulayo, futhi uma isiqondiso sokushintsha sifana nalokho okulindeleke ithiyori yezomnotho: ukubuyekezwa okuphansi okuvela ekuxhumaneni kwe-FOMC, izibuyekezo ezikhuphukayo ezivela ezindabeni zamanani aphezulu kaphethiloli. Isheke lama-ejenti ethu okwenziwa ukuthi ingabe izibuyekezo zawo zihlukanisa ngendlela efanayo abaphendula ngayo abantu.
Sakhe ama-synthetic personas angu-30,000 ngezibalo zabantu ezisuselwe kuCensus, futhi salinganisela umphumela wokwelapha omaphakathi ku-LLM ngayinye kwamathathu, okuhlanganisa nalawo angafundiwe. Isheke lokuqala liphathelene nezinto eziza kuqala ngokwazo: ama-ejenti alindeleke e-inflation abika ngaphambi kokuba abone noma yiluphi ulwazi. Umfanekiso wesi-3 uhlela ukuchezuka kwencazelo nokujwayelekile kwalokhu okubalulekile kuwo wonke amaqoqo amancane esibalo sabantu kubhentshimakhi yabantu kanye nama-LLM amathathu. Imodeli eyodwa engafundi (i-Llama-GA) isondela ekuhlanganiseni komuntu kukho kokubili ileveli nokuhlakazeka. Ngenkathi enye indlela yokungafundi isebenza (GA), enye ayisebenzanga (NPO). Ngakho-ke ukungafundi kungase kungabi ikhambi elilingana nosayizi owodwa.
Umfanekiso 3 Izilinganiso Zemodeli Ye-Inflation Ecatshangwayo
Qaphela: Iphaneli ngayinye idweba ngeqembu elincane lesibalo sabantu kubhentshimakhi yomuntu (Coibion et al., 2022), isisekelo se-Llama-3, kanye nezinhlobonhlobo zayo ezimbili ezingafundiwe (GA, NPO). Umugqa wedeshi uphawula inani elithi “Konke” lomuntu. Uhlangothi lwesobunxele: I-Llama-3 ne-Llama-NPO empeleni ziyizicaba kuzo zonke izici zezibalo zabantu; I-Llama-GA ilandelela izinga lomuntu ngokwesilinganiso kodwa ayikhiqizi kabusha ukuhleleka kwabantu (isb ukubikezela incazelo ephezulu kakhulu yokuthi “ikolishi noma ngaphezulu” kanye “ne-Inc T3,” ngokuphambene nephethini yomuntu). Uhlangothi olungakwesokudla: imodeli ye-GA engafundiwe ibuyisela iningi lokuhlakazeka eligoqwe imodeli yesisekelo.
Ukuhlola okulandelayo kumayelana nokuthi izinto ezandulelayo zibuyekezwa kanjani ngemva kokwelashwa kolwazi. Kumamodeli ayisisekelo e-Llama-3 kanye ne-Llama-NPO, ukubuyekezwa kuyefana ngokuyinhloko kukho konke ukwelashwa futhi amamodeli awawubhalisi nhlobo umphumela wokwelapha. I-Llama-GA iyona kuphela lapho ukwelashwa kwehlukana khona, futhi ngaphakathi kweqembu layo elikhulu lama-ejenti (80% yesampula) izindlela zokwelapha ezine zenqubomgomo yemali (ukwehla kwamandla emali okwedlule, i-Fed target, isibikezelo se-FOMC, isitatimende se-FOMC) kukhiqiza ukubuyekezwa okungalungile nokubalulekile kwesibonakaliso esifanayo kanye nobukhulu obunzima njengabaphenduli babantu ku-Coibion et al.
Yini okufanele uyithathe kulokhu
Kubacwaningi nodokotela abathatha isinqumo sokuthi basebenzise ama-LLM ukwenza inhlolovo, isifinyezo sithi:
- Ama-LLM awakwazi ukulingisa abantu abahlukene. Izinhlolovo ezilingisayo zehlela kumenzeli oyedwa ophendula umbuzo ofanayo izikhathi eziyizinkulungwane, eshaya into eseduze kakhulu nencazelo njalo, kwesinye isikhathi kufika ezindaweni ezine zamadesimali.
- Ukungafundi okuhlosiwe kubuyisela iningi lokuhlakazwa kanye nesabelo esihloniphekile semiphumela yokwelapha ku-RCT nabantu abaphendulayo. Nokho, izindlela zokungafundi zifinyelela amazinga ahlukene empumelelo.
- Igebe phakathi kokunemba kwencazelo kanye nokunemba kokusabalalisa likhulu ngokwanele ukuthi noma yiliphi iphepha elisebenzisa abaphenduli bokwenziwa kufanele libike okwesibili.
Umsebenzi wesikhathi esizayo kufanele uphathe ukunemba kokusabalalisa kanye nokuvuza kwedatha njengezithiyo ezihlanganyelwe esikhundleni sokukhathazeka kwesibili. Inqubekelaphambili izoncika ezindleleni ezibalela kukho kokubili lokho amamodeli akwaziyo nokuthi imiphumela yawo ihlolwa kanjani, kugxilwe kakhulu ekuhlakazweni, emisileni, nasekuthuthukisweni kwezinkolelo kunezilinganiso zodwa.
Izikhombo
U-Coibion, O., Y. Gorodnichenko, no-M. Weber (2022). Ukuxhumana kwenqubomgomo yezimali kanye nemithelela yakho ekulindelweni kokwehla kwamandla emali ezindlini. Ijenali Yomnotho Wezepolitiki 130(6), 1537–1584.
U-Dalloul, A., Pfeifer, M. (2026). Ingabe Ama-LLM Angakwazi Ukulingisa Izinhlolovo Zasekhaya?: Kusukela Kubameli Abamele Ukuya Ekusabalazweni Kwabantu. I-SSRN ukuphrinta ngaphambili. Xhumanisa nephepha lokusebenza
U-Zarifhonarvar, A. (2026). Ukukhiqiza okulindelekile kukwehla kwamandla emali ngamamodeli ezilimi ezinkulu. JUmlando wethu we-Monetary Economics 157103859
Idatha yokuphindaphinda
U-Dalloul, A., Pfeifer, M. (2026). Idatha yokuphindaphinda ye-: “Ingabe ama-LLMs Angalingisa Izinhlolovo Zasekhaya?: Kusukela Kubenzeli Abamele Ukuya Ekusabalaliseni Kwabantu”, i-Harvard Dataverse, V1.
