Kiyasin Tsarin Bakanci na Bayesian wanda ba na ƙa'ida ba don Jerin Lokaci tare da Canjin Ƙarfin Lokaci-Lokaci

1. Gabatarwa

Ƙirar daidaitaccen yanayin kuskure yana da mahimmanci a cikin nazarin jerin lokaci, musamman ga bayanan tattalin arziki da na kuɗi inda heteroskedasticity ya zama ruwan dare. Hanyoyin gargajiya sau da yawa suna sanya ƙayyadaddun tsarin ƙa'idodi akan kuskuren autocovariance, suna haifar da haɗarin kuskuren samfurin. Wannan takarda tana ba da shawarar hanyar Bayesian wacce ba ta ƙa'ida ba don kimanta girman bakanci na kuskuren autocovariance, tana magance yanayin ƙarfin da ke canzawa da na lokaci-lokaci. Hanyar tana kaucewa matsalar zaɓin bandwidth mai wahala da ke cikin hanyoyin da ba na ƙa'ida ba ta hanyar aiki a cikin yankin mitar tare da fifikon tsarin Gaussian.

2. Hanyar Aiki

2.1 Tsarin Samfurin

Babban samfurin shine tsarin koma baya: $y = X\beta + \epsilon$, inda $\epsilon_t = \sigma_{\epsilon, t} e_t$. A nan, $e_t$ tsarin Gaussian ne mai rauni tare da aikin autocorrelation $\gamma(\cdot)$, kuma $\sigma^2_{\epsilon, t}$ yana wakiltar ƙarfin canjin lokaci. An mayar da hankali kan girman bakanci $\lambda(\cdot)$ na $e_t$.

2.2 Kiyasin Bakanci na Bayesian wanda ba na ƙa'ida ba

Biyon Dey da sauransu (2018), an sanya fifikon tsarin Gaussian akan girman bakanci da aka canza zuwa log $\log \lambda(\omega)$. Wannan fifikon yana da sassauƙa kuma yana guje wa ƙayyadaddun zato na ƙa'idodi. Ana ci gaba da ƙididdigewa ta hanyar tsarin Bayesian mai matakai, yana haifar da rarraba bayanai na $\lambda(\cdot)$, $\beta$, da sigogin ƙarfi.

2.3 Ƙirar Ƙarfin Canjin Lokaci

An ƙirƙira ƙarfin log $\log \sigma^2_{\epsilon, t}$ ta amfani da ayyukan tushen B-spline, yana ba da wakilci mai sassauƙa na bambancin canji akan lokaci. Wannan ya faɗaɗa aikin Dey da sauransu (2018) ta hanyar ƙirar heteroskedasticity a fili.

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Babban ƙirƙira yana cikin ƙayyadaddun fifikon haɗin gwiwa da amfani da kusancin yiwuwa a cikin yankin mitar. An ƙirƙira girman bakanci kamar haka: $$\lambda(\omega) = \exp(f(\omega)), \quad f \sim \mathcal{GP}(\mu(\cdot), K(\cdot, \cdot))$$ inda $\mathcal{GP}$ ke nufin tsarin Gaussian tare da aikin ma'ana $\mu$ da kernel covariance $K$. An yi amfani da kusancin yiwuwar Whittle don ingantaccen lissafi: $$p(I(\omega_j) | \lambda(\omega_j)) \approx \frac{1}{\lambda(\omega_j)} \exp\left(-\frac{I(\omega_j)}{\lambda(\omega_j)}\right)$$ inda $I(\omega_j)$ shine periodogram a mitar $\omega_j$. Don ƙarfin canjin lokaci, samfurin B-spline shine: $\log \sigma^2_t = \sum_{k=1}^K \theta_k B_k(t)$, tare da fifiko akan ƙididdiga $\theta_k$.

4. Sakamakon Gwaji & Bincike

4.1 Nazarin Kwaikwayo

An tabbatar da hanyar akan bayanan da aka kwaikwaya tare da sanannun tsarin autocorrelation (misali, hanyoyin ARMA) da ƙarfin stochastic. Mai ƙididdigewa na Bayesian wanda ba na ƙa'ida ba ya samu nasarar dawo da ainihin girman bakanci da hanyoyin ƙarfi, tare da ƙungiyoyin aminci na baya sun rufe ainihin ayyukan. Ya nuna ƙarfin juriya ga kuskure idan aka kwatanta da madadin ƙa'idodi kamar samfuran AR da aka kuskura.

4.2 Aikace-aikacen Hasashen Canjin Kuɗi

Sakamako na Farko: An yi amfani da samfurin da aka ba da shawara don hasashen manyan farashin musayar kuɗi (misali, USD/EUR, USD/JPY). An kimanta aikin hasashensa da samfuran ma'auni ciki har da Tafiya Bazuwar (RW), ARIMA, da samfuran GARCH.

Aikin Hasashe (RMSE)

Samfurin Bayesian da aka Ba da Shawara: 0.0124
Tafiya Bazuwar: 0.0151
GARCH(1,1): 0.0138
ARIMA(1,1,1): 0.0142

Lura: Ƙananan Kuskuren Tushen Ma'anar Murabba'i (RMSE) yana nuna mafi kyawun daidaiton hasashe.

Samfurin da aka ba da shawara ya sami RMSE mafi ƙasa, yana nuna fa'idarsa ta gasa. Ƙarfin samfurin na kama tsarin dogaro (ta hanyar girman bakanci) da heteroskedasticity ya ba da gudummawar mafi daidaitaccen hasashe na maki da yawa fiye da RW mai tsauri ko daidaitattun samfuran GARCH.

5. Tsarin Bincike: Fahimta ta Asali & Zargi

Fahimta ta Asali: Ainihin gudummawar wannan takarda ba wani samfurin Bayesian kawai ba ce; juyawa ce mai dabarci daga yaƙi da "la'anar girma" a cikin hanyoyin da ba na ƙa'ida ba na yankin lokaci zuwa amfani da "albarkar santsi" a cikin yankin mitar. Ta hanyar sanya fifikon Tsarin Gaussian kai tsaye akan girman bakanci na log, marubutan sun kaucewa zaɓin bandwidth na masu ƙididdigewa na kernel da sanannen wahala. Wannan yayi kama da falsafar da ke bayan samfuran samarwa mai zurfi masu nasara kamar CycleGAN (Zhu et al., 2017), wanda ke amfani da zagayowar adawa don koyon taswira ba tare da bayanan haɗin gwiwa ba—dukkan takardun sun warware matsala mai wahala ta hanyar sake tsara ta a cikin sarari mafi sauƙi (mita don jerin lokaci, zagayowar hoto don fassarar).

Kwararar Ma'ana: Hujja tana da ƙarfi: 1) Zato na ƙa'idodi akan kurakurai yana da rauni kuma yana haifar da kuskure (gaskiya, duba ɗimbin wallafe-wallafen kan rashin isasshen samfurin GARCH). 2) Hanyoyin da ba na ƙa'ida ba na gargajiya suna da aibi mai mutuwa (zaɓin bandwidth). 3) Je Bayesian kuma je yankin mitar inda fifikon GP yake aiki azaman mai sassautawa ta atomatik. 4) Kar a manta da ƙarfi—ƙirƙira shi ma cikin sassauƙa tare da splines. 5) Tabbatar yana aiki akan ma'auni mafi wahala a cikin kuɗi: doke Tafiya Bazuwar a cikin forex.

Ƙarfi & Kurakurai: Ƙarfi: Haɗin hanyar aiki yana da wayo. Haɗa fifikon GP don spectra tare da splines don ƙarfi shine bugun guda biyu mai ƙarfi don jerin lokaci na kuɗi. Nasara ta zahiri akan RW tana da ma'ana; kamar yadda aikin Meese da Rogoff (1983) ya kafa, wannan babban ma'auni ne. Kasancewar lambar a kan GitHub (junpeea) babban fa'ida ne don sake yin samfuri. Kurakurai: Farashin lissafi shine giwa a cikin daki. MCMC don fifikon GP akan spectra, haɗe tare da ƙididdigar ƙarfi, yana da nauyi. Takardar ba ta yi magana ba game da ƙididdiga na zamani na bambance-bambance ko ƙarancin GP don auna wannan. Bugu da ƙari, zaɓin B-splines don ƙarfi, duk da cewa yana da sassauƙa, ba shi da fahimta fiye da samfuran ƙarfin stochastic tare da jihohin ɓoye. Kwatancen hasashe, duk da cewa yana da kyau, ya kamata ya haɗa da ƙarin ma'auni na zamani kamar LSTMs na koyon zurfi ko samfuran tushen Transformer, waɗanda ke zama daidaitattun a cikin kuɗi mai yawan mitar (kamar yadda aka gani a cikin albarkatu daga Cibiyar Nazarin Manufofin Tattalin Arziki ta Stanford).

Fahimta Mai Aiki: Ga masu ƙididdiga da masana tattalin arziki: Wannan tsari ne don gina samfuran hasashe masu ƙarfi, masu tsarin rabin tsari. Abin da za a ɗauka shine daina tilasta tsarin kurakurai cikin akwatunan ARMA ko GARCH. Aiwatar da hanyar GP ta bakanci don kowane samfurin inda binciken ragowar ya nuna hadadden autocorrelation. Ga masu bincike da ake amfani da su, yi amfani da wannan a matsayin madadi mafi girma ga daidaitattun kurakurai na Newey-West lokacin da dogaro ba a sani ba. Gaba yana cikin samfuran gauraye: saka wannan kayan aikin kuskure wanda ba na ƙa'ida ba cikin manyan VARs na tsarin ko tsarin yin hasashe na yanzu. Babbar dama tana cikin haɗa wannan hanyar GP ta yankin mitar tare da Hamiltonian Monte Carlo (HMC) a cikin Stan ko PyMC don aiwatarwa mai amfani, mai iya aunawa.

6. Misalin Tsarin Bincike

Yanayi: Nazarin kudaden yau da kullun na cryptocurrency (misali, Bitcoin) don hasashen ƙarfinsa da tsarin dogaro, wanda aka san yana da hadadden kuma ba na tsayayye ba.

Matakan Aikace-aikacen Tsarin:

Ƙayyadaddun Samfurin: Ayyana samfurin ma'ana mai sauƙi (misali, ma'anar akai-akai ko koma baya akan kudaden da suka gabata). An mayar da hankali ga kalmar kuskure $\epsilon_t$.
Fifiko na Bayesian:
- Girman Bakanci ($\lambda(\omega)$): Sanya fifikon Tsarin Gaussian tare da kernel Matérn akan $\log \lambda(\omega)$ don kama dogaro mai santsi amma mai yuwuwar dogon ƙwaƙwalwar ajiya.
- Ƙarfin Canjin Lokaci ($\sigma^2_t$): Yi amfani da B-spline mai siffar cubic tare da ƙulli 20-30 akan jerin lokaci don ƙirar $\log \sigma^2_t$. Sanya fifikon daidaitawa (misali, tafiya bazuwar) ga ƙididdiga na spline don hana wuce gona da iri.
- Ƙididdiga na Koma Baya ($\beta$): Yi amfani da daidaitattun fifiko masu ba da labari mara ƙarfi (misali, Al'ada tare da babban bambanci).
Ƙididdiga: Yi amfani da samfurin Markov Chain Monte Carlo (MCMC) (misali, ta hanyar Stan ko samfurin Gibbs na al'ada) don samun rarraba bayanai na haɗin gwiwa na duk sigogi: $p(\lambda(\cdot), \sigma^2_{1:T}, \beta | \text{bayanan})$.
Fitowa & Fassara:
- Bincika ma'anar bayanai na $\lambda(\omega)$ don gano manyan mitoci na dogaro (misali, zagayowar gajeren lokaci da na dogon lokaci).
- Bincika yanayin bayanai na $\sigma^2_t$ don gano lokutan babban ƙarfi da ƙananan ƙarfi (misali, daidai da abubuwan kasuwa).
- Ƙirƙiri hasashe ta hanyar kwaikwayon hanyoyin gaba daga rarraba hasashe na baya, haɗa da ƙididdigar dogaro da ƙarfi.

Wannan tsarin yana ba da cikakken bayanin yiwuwar yanayin jerin ba tare da ɗaukar takamaiman siffar ARMA-GARCH ba, yana mai da shi ya dace da siffofin musamman na kasuwannin crypto.

7. Hasashen Aikace-aikace & Hanyoyin Gaba

Aikace-aikace na Nan take:

Hasashen Tattalin Arziki na Macro: Haɓaka samfuran yin hasashe na yanzu don GDP, hauhawar farashin kayayyaki, ko fihirisar damuwa ta kuɗi ta hanyar samar da mafi kyawun tsarin kuskure don samfuran masu yawan masu hasashe.
Gudanar da Haɗari: Inganta lissafin Ƙimar Haɗari (VaR) da Tsammanin Gajeriyar (ES) don fayil ɗin kadarori ta hanyar ƙirar haɗin gwiwar haɗin gwiwa da ƙarfin gefe na kudaden da aka dawo da su daidai.
Lissafin Tattalin Arziki na Yanayi: Ƙirar dogon ƙwaƙwalwar ajiya da heteroskedasticity a cikin jerin zafin jiki ko hayaƙin carbon, inda samfuran ƙa'idodi na gargajiya zasu iya kasawa.

Hanyoyin Bincike na Gaba:

Ma'aunin Lissafi: Haɗa ƙididdiga na ƙarancin tsarin Gaussian ko ƙididdiga na bambance-bambance don sarrafa jerin lokaci masu yawan mitar ko dogon lokaci.
Faɗaɗa Multivariate: Haɓaka fifikon GP na matrix-variate don girman bakanci na tsarin kuskure na vector, mai mahimmanci don nazarin fayil.
Haɗawa tare da Koyon Zurfi: Amfani da ƙididdigar girman bakanci azaman siffa ko mai daidaitawa a cikin samfuran jerin lokaci na tushen jijiyoyi (misali, Masu Fassara Haɗin Lokaci).
Ƙididdiga na Ainihin Lokaci: Haɓaka sigogin Monte Carlo na jeri (tace barbashi) na hanyar don hasashe kan layi da sa ido.
Ƙididdiga na Dalili: Yin amfani da samfurin kuskure mai sassauƙa a cikin tsarin sakamako mai yuwuwa don jerin lokaci don samun mafi ƙarfin daidaitattun kurakurai don tasirin magani.

Hanyar ta kafa tushe don sabon nau'in samfuran jerin lokaci na "agnostic" waɗanda suke da ƙarfi ga kuskure, hanyar da masu bincike a Hukumar Binciken Tattalin Arziki ta Ƙasa (NBER) suka ba da shawarar sosai don tattalin arzikin zahiri.

8. Nassoshi

Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica, 50(4), 987-1007.
Kim, K., & Kim, K. (2016). A note on the stationarity of GARCH-type models with time-varying parameters. Economics Letters, 149, 30-33.
Dey, D., Kim, K., & Roy, A. (2018). Bayesian nonparametric estimation of spectral density for time series. Journal of Econometrics, 204(2), 145-158.
Kim, K. (2011). Hierarchical Bayesian analysis of structural instability in macroeconomic time series. Studies in Nonlinear Dynamics & Econometrics, 15(4).
Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired image-to-image translation using cycle-consistent adversarial networks. Proceedings of the IEEE international conference on computer vision (pp. 2223-2232).
Meese, R. A., & Rogoff, K. (1983). Empirical exchange rate models of the seventies: Do they fit out of sample? Journal of international economics, 14(1-2), 3-24.
Whittle, P. (1953). Estimation and information in stationary time series. Arkiv för Matematik, 2(5), 423-434.