CC BY
17Uphhumm^mU UhjpnUmj^U gmUghpp qUhp^. Unqhim^nptfmU qnpbpUpmgniU
-0-linujmh U. b'Z (Zmjmummh, bplimh) atevossyan@yahoo.com
4^nnpn2 pmnhp' Um^muUnijg mphmhurnm^mU UhjpnUmj^U gmUghp, Unqhj^ qhpmhmpUmphgniU, qUhp^ Unqhjm^npnLU
Искусственные нейронные сети в процессе анализа цены
Тевосян А. ЕГУ (Армения, Ереван) atevossyan@yahoo.com
Резюме: В статье рассматривается возможность использования нейросетей Feed Forward для моделирования цен. В статье также представлены вопросы модельных перестановок в процессе машинного обучения. Рассматриваются три этапа генерации моделей в машинном обучении: обучение, валидация (калибровка) и тестирование (оценка), которые помогают избежать переустановок моделей.
Ключевые слова: искусственные нейронные сети, переустанови моделей, прогнозирование инфляции
Artificial neural networks in the process of price analysis
Tevosyan A.
International Scientific-Educational Center, NAS, RA (Armenia, Yerevan)
atevossyan@yahoo.com
Abstract: The article discusses the possibility of using Feed Forward Neural Networks for price modeling. The article also presents model overfitting issues in machine learning models. It discusses the three stages of model generation in machine learning: training (model fit), validation (calibration) and testing (evaluation), which help avoid overfitting. Keywords: artificial neural networks, model overfitting, inflation forecasting
UhphUmUhp^ nLuntgntUU ^phU^g Uhp^m-jmgUntU t hmUm^mpq^Uhp^U ^UpUm^mp ^hp-цп^ ntunLgmUh^nL q^mntpjntU h^UU^mb ^pm-^mU m2^mph^ фn^шqqhgnLp]nLUUhp^ U Umpq^mj^U ^mp^hjm^hp^ ntuntUUmupUmU, uhpmUmU, ^p^Uop^Um^UmU, U np mUhUm-^mpUnpU t npm^m^mU pmphjm^ntUp U o^m^U^mimgntUp m^mhn^nq mjqnppUhUp^ ^pm: UhphUmUhp^ ntuntgntUp ^mmmp^ntU t pmqUmp^ Unqh^Uhpn^: UnLjU hnq^mbntU UhUp ^nuhUp UhphUmUhp^ ntuntgUmU phpUu mUhUmhqnp UnmhgUmU UhjpnUmj^U gmUghpp Umu^U: LhjpnUmj^U gmUghpp u^qpUmq&^mb hU hqhj_ Umpq^mj^U UmmbnqnLpjmU ^p^Um-^mm^hpn^: LhjpnUmj^U gmUghpp hmUq^um-UntU hU ^UpUmntunLgmU^nq hmUm^mpqhp: bUpUmntuntgUmU pUpmgpntU, nLUntUUmu^ph-jn^ U uhpmhjn^ pm^m^mU^U Uh& Ьш^шф mhqh^mm^nLpjntU, gmUghpp ^mpnqmUntU hU pmgmhmjmhL Untmpmj^U U hjpmj^U m^mjUhp^ U^gU hqmb pmpU^mb mUhUmpmpq U hm^m^
UmU pmpU^mb ^m^^mbntpjntUUhpp, Цштш-phj_ pUqhmUpmgntUUhp U qUmhmmntUUhp: ZhUg UhjpnUmj^U gmUghpp UnqhjUhp^ U^gngn^ t hUmpm^np hqh]_ h^UUhj_ mphhumm^mU рш-Um^mUntpjmU qшqшфшpp:
Uphhumm^mU UhjpnUmj^U gmUghpp rnh-um^p, npp mnm^hjm^hu oqmmqnpb^ntU t qUhp^ Um^mpq^ Unqhjm^npUmU hmUmp, Um^muUntjg mphmhumm^mU UhjpnUmj^U gmUghpU hU (Feed Forward Neural Networks Цш^ FFANN), npnUp oqmmqnpbhjn^ pmqUm2hpm pU^m^Uhp, ^ hU pU^mjh
шUuшhUшUшфшЦ pmUm^ 2mpn^Um^m^mU фnLUЦg^шUhp U Unrnmp^h^ qpmUp ^hpmbhjn^ pUqhmUpm^mU mpqjntUpUhp^:
LhjpnUmj^U gmUghpp ^mpnq hU U^mpm-qp^h np^hu ^um^mpmUhmp^ UnqhjUhp: U^ ^nqU^g, Unqh^ ^mnntg^mbpp U ^m^hpp Um^mU2^ntU hU gmUgmj^U ^mnntg^m&p^ U^gngn^, UjnLu ^nqU^g, um^mjU, gmUghpp phUp hU np^ntU, ph Untmpmqp^mb m^mj]_-
Регион u Mup, 2019, № 6
Uhpp ^mjdmUUhpnLd n p Im^hpU hU ImpUnp
a
li n p Ijnliljphui iquipuiifhuiphpp:
Qbm^mslhp 1: bmJumuhnLjg wpihhuinwliwh hhjpnhmjpiii gmUghpp (Feed Forward Neural Networks
imiiFFANN) qpm$>pilmlmh UhplmjmgnLiIp:
SmUghpp nLUmlnLpjnLUp mnm^h]m^hu pmphim^nLd t, hpp Um^uUnLjg s^jmiUhpp pmUmlp m^nLd t:
UnmgUmjpU ^nL^nL^ UhpUmb^mb Uhpp pmji-mn-pmji ^n^qnpbmlgnLpjmU dhg hU UsUnLd, U jnLpmpmUynLp hmUqnLjg (Qbm-^mmlhp 1, qhqpU hmUqnLjgUhpp) pphUpg Uhp-ImjmgUnLd t hmdmlg^mb ^n^qnpbnLUhnL-pjmU mqqmU2mU UhlUhlnL n^ qbmjpU $nLUl-gpmUhpp dpgngn^: UjUnLhhsU, hmgnpq ^nL^nL^, ^n^qnpbmlgnLpjnLUp m^hip t ^npmUnLd U pmphim^nLd, hpp ulunLd hU A.Um^np^hi UhpUmb^mb s^jmiUhpp l?pnUhpp U hmpmphpmlmU ^mpmdhsphpp: Uju ^nL-inLtf qnpbmb^nLd t n^ pmgmhmjs ^n^qnp-bmlgnLpjmU 2hpsp (Q-bm^mslhp 1, ^mpqm-qnLjU hmUqnLjgUhpp): $mpU^mb hmUqnLjgUhpp pphUgpg UhplmjmgUnL^ hU pmphim^dmU imtf mlsp^mgdmU ^nLUlgpmUhp (activation function), npnUp uhprnh^n^ gmUgmjpU ^n^lm-^mlg^mbnLpmU hmslmUp2Uhpp U hmpdmp-^h^n^ pmqdmpp^_ ^mppmqbmjpU qpuUnpnLd-Uhpp U^mqmqnLjU 2hqnLdUhpn^ UdmUmlnLd hU ^hpgUmlmU mpqjnLUpp, mju uh^pnLd qUm^p (Q-bm^mslhp 1, Im^nLjs hmUqnLjgp): CUqhmUpm^hu, n^ qbmjpU hmpmphpnLpjnLU-Uhpp dnqhim^npnq mlsp^mgdmU ^nLUlgpm-Uhpp nLUhU inqpuspl imd upqdnpq shup: Zm^m^ UmU, hpp imppp t LpUnLd hplnL imtf m^hip qmuhpp qmumlmpqmjpU UnqhiUhp qmpqmgUh^nL (classes of classification models) oqsmqnpb^nLd t $ppqdmUp un^pdmpu
(saftmax) ^nLUlpgpmU ($ppqdmU U mjjnp, 2009):
shumUhip t Qbm^mslhp 1-pg, W1 l?pnUhpp dmsppgp UhpUmb^mb s^jmi-Uhpp hmJmlgnL^ t gmUgp pmpU^mb ^n^qnpbmlgnLpjm 2hpsp hhs, npshqpg U A.Um^pn^nLj hU hmslmUp2Uhpp, npnUp W2 l2pnUhpp "dmsppgp dpgngn^ AUm^npnLd t UhpbpU ^n^n^mlmUp qUm^p: ZmUqnLjghpp mlsp^mgpnU ^nLUlgpmjp qnpbmlpgUhpp hhUg l?pnUhpp dmsppgp smpphpU hU: UjupUpU, i-pq hmUqnLjgp k-pq 2hpsp hmdmp mlsp^mgpmU ^nLUlgpmjp UmspmjpU mpdhp t i-pq hmUqnLjgp (k-1) 2hpsp l?pnUhpp tfmsppgp qnpbmlpgp: Qbm^mslhp 1-p ^mpq puiliuiAllui] Jill uihqpp IjiJitiJi htonlijuqp
Hpshq X*W pphUpg U UhplmjmgUnLtf t X hmslmUp2p U l?pnUhpp tfmsppgp mpsmqp-jm^p: ^2hUp, np ^nq^um^l nhqphu^mU hUmpm-^npnLpjnLU t rnmi^u UhpUmbh]_ mU^n^nq qnpbml^g (2hqnL^) gmUlmgmb 2hpmnLtf:
Un^npmpmp mqms ^mpmJhmphp^ pm-Umlp npn2^nL^ t pmpU^mb 2hpshp^ pmUm-ln^ U jnLpmpmU^jnLp 2hprnnLj hmUqnLjgUhpp p^n^: SmUghpU, npnUp nLUhUnL^ hU m^h^p pmU Jhl pmpU^mb 2hps, In^^nL^ hU ^npp, U hhUg mjushqpg tL ^npp nLunLgnL^ (deep learning) (Lh£nLU U mjLnp (2015) 1: QnLq-^hLn U mjLnp (2016)2): QUmjmb np un^npmpmp mlUlmi^nLd, np mqms ^mpmdhmphpp m^p hhs dhlmhq dnqhip m^hip im^ npqjnLUp ism, dhb t hm^mUmlmUnLpjnLUp, np um iphpp UmU dnqh^ qhphmpdmpdmphgdmU (overfitting): QhpmhmpdmphgnLdp mphhusmlmU pmUm-ImUnLpjmU U JhphUmUhpp nLunLgdmU dhg mdhUmppmlmU U ^smUqm^np ^UqppUhppg t: Cms qh^phpnLtf miqnpppdp ^m^pg im^ t hmpdmp^nLd s^jmiUhppU mnm^him^hu IhUspnUmUmin^ dnqhip mqdnLlp (noise), mji n^ ph mqqmU2mUUhp ^pm: Um 2ms hm^m^
1 LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553):436-444.
2 Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press.
^msmhnri t mjU qh^nri hpp sljmLUhpp umlml hU, ^ul inqhLp ¿m^g mnm^h pmpq hmpnrus hmUqnLjgUhpnl: UpqjnLUpnrJ nrUh-Unri hUp gmUgmj^U mLqnp^pd, npp phU lm-smphLm^hu hmpimplh t Unp s^jmLUhp^U, pmjg pnLjL t U ¿^ ImpnqmUnrJ lmU-
^mshunrdUhp ImsmphL hmUqhgUhLnl gm&p npml^ U ^pmlmUnrpjmUp ¿hmJmqmmmu-^mUnq mpqjnrUp^: Un^npmpmp, mju qh^ph-pnii, hmpl t JhphUmj^ nLunignrip u^uh[ hmihimsmpmp ^mpq U inpp pmjLhpnl, in-qhLP Inp&mplhL ImlhpmgimU hmimp mnmU&pmgpmfc sljmLUhpp^ ipm U ImsmphL ^m^mA-U ImlhpmgnLiUhp (validation / cross-validation): Unlnpmpmp, dhqhLUhp^ qhpmhmp-imphgimU ^Uqp^g ^nrum^hLnr hmimp, ^np-hnrpq t sp^nti hhsUhL dnqhLUhp^ InpAmp-limU U ImlhpmgimU hphp pmjLhp^U: Upm Um^m^mjdmU t mdpnqgm^mU p^mj^U s^jmL-Uhp^ pmgmUnrip hphp dmu^ nrunrgdmU, ImlhpmgimU U usnrqimU ^ummqnrjUu ^mh^mUhLn^ ^mmmhm^mU pUspmUp^ ulqpnrUpp pnLnp hphp s^mj^U hUpm^iphp^ hmimp: ¿mpl t U2hL um^mjU, np pnLnp hphp ^Jphpnri pm^mpmp pmUml^ q^smplnrdUhp nrUhUmLnr hmimp mdpnqgm^mU p^mj^U m^jmLUhp^ hmlmpm&nrU/pmqmU ^hsp t L^U^ pm^mpmp dh&:
UnqhLUhp^ UdmUms^ ihpm^mspmus-imU hnm^nqd gfrlp inqh^ Ahrm^npnrdp, hh-smqm im^hpmgdp hr inp&mplnrdp d^mjh impnq t m^mhn^hL dnqh^ ^ummhhj^ pUqhm-pmgUnq hmm^nLpjnrUUhp:
Uju^unl dnqh^ o^s^dmL q^qmjUp im^mfc t m^jmLUhp^ 2mhdmpmU^ ¿m^^g,
hmpg^ pUnrjp^g U 2ms hm^m^ dnqhLm^npnq^ iJinpAfig U qjiuihjjiplihplig:
^bm^msqtp 2: Uhphhmhhpji ntuntgilmh
tlnqhihhp]i hphp ipnLihpp u]uhimin]il Uhplmjm-gnLiI' uLunLguLii(inqhrfi hmilmvqmrnmuJumhhgnLiIp lnl]Jmlhhp]lh), ilmilhpmgml(umuLquLi) hL iJmp&mplnLi (qhmhminnLl):
Oqwmqnp&^mb qpmlmUntpjmh gmh^
1. LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553):436-444.
2. Glass E., Big data in central banks: 2016 survey, Central Banking Journal 27 (2) November 2016, last modified November 7, 2016, www.centralbanking.com/2474825 2.
3. Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press.
4. Chiranjit Chakraborty and Andreas Joseph, Machine learning at central banks, September 2017
5. Thomas R. Cook and Aaron Smalter Hall, Macroeconomic Indicator Forecasting with Deep Neural Networks, September 2017.
6. Koop G., Onorante L., Macroeconomic Nowcasting Using Google Probabilities, August 2013.
7. Bholat, D. (2015). Big data and central banks. Bank of England Quarterly Bulletin, 55(1):86-93.
8. Einav, L. and Levin, J. D. (2013). The data revolution and economic analysis. Working Paper 19035, National Bureau of Economic Research.
