Title Identifikacija strukture neuronske mreže u mjerenju očekivane inflacije Author Tea Poklepović Mentor Josip Arnerić (mentor)Committee member Josip Arnerić (član povjerenstva)Granter University of Split Defense date and country 2017-10, Croatia Scientific / art field, SOCIAL SCIENCES Universal decimal classificationUDC ) 330 - Economics. Economic science Abstract Neuronske mreže (NM) je prikladno koristiti u analizi vremenskih nizova u uvjetima narušenih pretpostavki, tj. u uvjetima „nenormalnosti“ i nelinearnosti. Svrha rada je istražiti nedostatke NM te predložiti načine kojima se ti nedostaci mogu otkloniti s ciljem identifikacije strukture NM koja će se najbolje prilagoditi inflacijskim očekivanjima. Polazi se od teorijskog modela koji uključuje karakteristike inflacije potražnje i inflacije troškova. Pri tom se koriste varijable tržišta rada,
... More financijske varijable, eksterni faktori te inflacija iz prethodnog razdoblja. Istraživanje se provodi na agregiranoj razini zemalja euro zone u periodu od siječnja 1999. do siječnja 2017. godine. Na temelju procijenjenih 90 jednosmjernih NM i 450 Jordanovih NM, koje se razlikuju u promjenjivim parametrima (broj iteracija, stopa učenja, interval početnih vrijednosti težina, broj skrivenih neurona, vrijednost težine kontekstne jedinice), dobiveni su pokazatelji prikladnosti za svaku NM (prosječna kvadratna pogreška - MSE i Akaikeov informacijski kriterij - AIC) koji se odnose na dva perioda: „unutar uzorka“ (engl. in the sample) i „izvan uzorka“ (engl. out of sample). Istovremeno rangiranje NM na poduzorcima „in-the-sample“ i „out-of-sample“ bilo prema MSE ili prema AIC ne dovodi do podudarnosti rangova i odabira prikladne NM jer najbolja NM na poduzorku „in-the-sample“, na temelju kriterija MSE i/ili AIC, često daje na poduzorku „out-of-sample“ visoke vrijednosti oba pokazatelja prikladnosti, i obrnuto. Stoga, da bi se došlo do najboljeg kompromisnog rješenja koristi se PROMETHEE metoda za odabir prikladne NM. Usporedbom „najbolje“ jednosmjerne – FNN(4,5,1) i „najbolje“ Jordanove NM – JNN(4,3,1) zaključuje se da je u približno jednakim uvjetima potrebno manje neurona u skrivenom sloju Jordanove NM (3) u odnosu na jednosmjernu (5), čime se potvrđuje da je Jordanova NM parsimonijska u odnosu na jednosmjernu, uz ne postojanje problema predeterminiranosti modela. Na taj način se dokazuje prva pomoćna hipoteza. Također, odabrana Jordanova NM ima bolju prediktivnu sposobnost u odnosu na jednosmjernu, čime se potvrđuje i druga pomoćna hipoteza istraživanja. Konačno se analizira konvergencija kamatnih stopa, uključivanjem inflatornih očekivanja procijenjenih Jordanovom NM u preformulirani oblik Taylorovog pravila. Less Abstract (english) Neural networks (NNs) are appropriate to use in time series analysis under conditions of the unfulfilled assumptions, i.e. under conditions of non-normality and nonlinearity. The purpose of the paper is to investigate the disadvantaged of NNs and to propose ways to address these shortcomings with the aim of identifying the NN structure that will best adapt to inflation expectations. The research is based on a theoretical model that includes the characteristics of demand-pull and
... More cost-push inflation, i.e. it uses labour market variable, financial variable, external factor and lagged inflation. The research is conducted at the aggregate level of euro area countries in the period from January 1999 to January 2017. Based on the estimated 90 feedforward NNs and 450 Jordan NNs, which differ in variable parameters (number of iterations, learning rate, initial weight value intervals, number of hidden neurons, weight value of the context unit), the model adequacy indicators for each NN (Mean Squared Error - MSE and Akaike Information Criterion - AIC) are calculated for two periods: “in-the-sample” and “out of sample”. Simultaneously ranking NNs on the “in-the-sample” and “out of sample” subsamples either according to MSE or AIC does not lead to the matching of rankings and the selection of a suitable NN because the “best” NN in the “in-the-sample”, based on MSE and/or AIC criteria, often has high “out of sample” values of both indicators, and vice versa. Therefore, in order to achieve the best compromise solution, a PROMETHEE method is used to select a suitable NN. By comparing the “best” feedforward NN - FNN(4,5,1) and the “best” Jordan NN - JNN(4,3,1) it is concluded that under approximately equal conditions less neurons in the hidden layer of Jordan NN (3) is required than in feedforward NN (5), confirming that Jordan NN is parsimonious compared to feedforward, without the existence of the overfitting problem. Thus, the first auxiliary hypothesis is proved. Moreover, the selected Jordan NN has a better predictive ability than the feedforward, which confirms second auxiliary hypothesis of the research. Finally, interest rate convergence is analysed, incorporating inflation expectations estimated by Jordan NN into the reformulated Taylor rule. Less Keywords
euro zona
jednosmjerna neuronska mreža
Jordanova neuronska mreža
konvergencija kamatnih stopa
očekivana inflacija
Taylorovo pravilo
Keywords (english)
euro zone
feedforward neural network
Jordan neural network
interest rate convergence
expected inflation
Taylor rule
Language croatian URN:NBN urn:nbn:hr:124:939653 Study programme Title: Postgraduate doctoral study programme in Economics and Business Type of resource Text File origin Born digital Access conditions Open access Terms of use Created on 2018-03-22 12:10:51
