1. Gabatarwa

Daidaicin hasashen farashin musayar kuɗi na RMB/USD kalubale ce mai mahimmanci a cikin kuɗaɗen duniya, wanda ke shafar ciniki, saka hannun jari, da manufofin kuɗi. Rashin kwanciyar hankali da kuma rikitarwa, da kuma sauye-sauyen da ba su da tsari a kasuwannin musayar kuɗi na duniya sun sa ƙirar tattalin arziki na gargajiya su zama marasa isa. Wannan binciken ya magance wannan gibi ta hanyar tantancewa cikin tsari na ƙirar koyon zurfi (DL) na ci gaba—ciki har da Ƙwaƙwalwar Ƙwaƙwalwa na Dogon Lokaci (LSTM), Cibiyoyin Jijiyoyin Kwakwalwa (CNN), da ƙirar Transformer—don hasashen farashin musayar kuɗi. Wani sabon abu mai mahimmanci shi ne haɗa fasahar AI masu bayyanawa (XAI), musamman Taswirar Ayyukan Aji Mai Nauyin Gradient (Grad-CAM), don bayyana yanke shawarar ƙirar da kuma gano mafi tasiri na fasalin tattalin arziki da na kuɗi.

2. Hanyoyi & Ƙirori

2.1 Bayanai & Injiniyan Fasali

Binciken ya yi amfani da cikakken bayanai na fasali 40 a cikin rukuni 6 don hasashen farashin RMB/USD. Rukunin fasali sun haɗa da:

  • Alamomin Tattalin Arziki: Haɓakar GDP, ƙimar hauhawar farashin kayayyaki (CPI, PPI), bambance-bambancen ƙimar riba.
  • Ciniki & Gudanar da Babban Kuɗi: Yawan ciniki tsakanin Sin da Amurka, ma'auni na asusun yanzu.
  • Farashin Musayar Kuɗi masu Alaka: Nau'ikan musayar kuɗi kamar EUR/RMB da USD/JPY.
  • Halin Kasuwa & Rashin Kwanciyar Hankali: Fihirisar rashin kwanciyar hankali da aka nuna, farashin kayayyaki (misali, mai).
  • Manufofin Kuɗi: Ƙimar manufofin bankin tsakiya da buƙatun ajiya.
  • Alamomin Fasaha: Matsakaicin motsi, na'urori masu motsi da aka samo daga bayanan farashin tarihi.

An yi amfani da tsari mai tsauri na zaɓin fasali don rage girma da kuma haskaka mafi yawan masu hasashe, tare da jaddada tushen tattalin arziki maimakon hayaniya.

2.2 Tsarin Koyon Zurfi

Binciken ya yi kwatankwacin ƙirori da yawa na zamani:

  • LSTM: Yana ɗaukar dogon lokaci na dogaro a cikin bayanai masu tsari.
  • CNN: Yana fitar da alamu da fasali na gida a cikin bayanan lokaci-lokaci.
  • Transformer: Yana amfani da hanyoyin kula da kai don auna mahimmanci na matakan lokaci daban-daban da fasali a duniya.
  • TSMixer: Ƙirar da aka tsara ta MLP don hasashen lokaci-lokaci, wanda ya fi sauran aiki a cikin wannan binciken. Yana amfani da sassa masu kauri a cikin lokaci da girma na fasali, yana ba da ƙira mai sauƙi amma mai inganci sosai don ɗaukar hulɗa masu rikitarwa.

2.3 Bayyanawa tare da Grad-CAM

Don ƙetare hanyar "akwatin baƙi", marubutan sun yi amfani da Grad-CAM, fasaha da aka ƙera don hangen nesa na kwamfuta (Selvaraju et al., 2017), don hasashen lokaci-lokaci. Grad-CAM yana samar da taswira mai zafi wanda ke nuna waɗanne fasalin shigarwa (kuma a waɗanne matakan lokaci) suka fi mahimmanci ga hasashen ƙirar. Wannan yana ba masu bincike damar tabbatarwa idan hankalin ƙirar ya yi daidai da fahimtar tattalin arziki—misali, ba da fifiko ga bayanan yawan ciniki a lokutan tashin hankali na ciniki.

3. Sakamakon Gwaji

3.1 Ma'aunin Aiki

An kimanta ƙirori ta amfani da ma'auni na yau da kullun: Matsakaicin Kuskure na Mutum (MAE), Tushen Matsakaicin Kuskure na Murabba'i (RMSE), da Matsakaicin Kuskure na Kashi na Mutum (MAPE).

Taƙaitaccen Aikin Ƙira (Bayanan Hasashe)

Mafi Kyawun Aiki (TSMixer): RMSE = 0.0052, MAPE = 0.68%

Transformer: RMSE = 0.0058, MAPE = 0.75%

LSTM: RMSE = 0.0061, MAPE = 0.80%

CNN: RMSE = 0.0065, MAPE = 0.85%

Lura: Takamaiman sakamakon lambobi suna misaltawa bisa labarin takarda na fifikon TSMixer.

3.2 Muhimman Bincike & Hotuna

Ƙirar TSMixer a koyaushe tana ba da mafi daidaitattun hasashe. Mafi mahimmanci, hotunan Grad-CAM sun bayyana fahimta mai amfani:

  • Mahimmancin Fasali: Ƙirar ta ba da muhimmanci sosai ga yawan ciniki na Sin-Amurka da farashin musayar EUR/RMB, yana tabbatar da mahimmancin haɗin ciniki na asali da cinikin musayar kuɗi.
  • Hankali na Lokaci: A lokutan matakan kasuwa masu tashin hankali (misali, bayan gyara na 2015, rikicin ciniki na 2018), hankalin ƙirar ya karkata sosai zuwa alamomin halin labarai da kwanakin sanarwar manufofi.
  • Bayanin Chati: Taswirar zafi na Grad-CAM na hasashe zai nuna hoton layi da yawa. Kowane layi yana wakiltar fasali (misali, Yawan_Ciniki, EUR_RMB). X-axis shine lokaci. Kwayoyin suna da launi daga shuɗi (ƙaramin mahimmanci) zuwa ja (babban mahimmanci). Muhimman lokuta suna nuna jajayen bandeji masu haske a cikin fasali na asali, a zahiri suna "bayyana" hasashen.

4. Nazari & Tattaunawa

4.1 Cikakken Fahimta & Tsarin Hankali

Cikakken Fahimta: Muhimmin gudummawar takarda ba kawai koyon zurfi yana aiki ba, amma cewa ƙira masu sauƙi, da aka tsara da kyau (TSMixer) na iya fi na rikitarwa (Transformer) aiki don takamaiman ayyukan hasashen kuɗi, musamman lokacin da aka haɗa su da ingantaccen injiniyan fasali da kayan aikin bayyanawa. Tsarin hankali yana da inganci: gano rikitarwar matsalar hasashe, gwada jerin ƙirar DL na zamani, sannan a yi amfani da XAI don tabbatarwa da fassara hankalin mai nasara. Wannan yana motsa fagen daga aikin hasashe kawai zuwa aikin da za a iya tantancewa.

4.2 Ƙarfafawa & Gazawar Muhimmi

Ƙarfafawa:

  • Haɗin XAI Mai Amfani: Yin amfani da Grad-CAM ga kuɗaɗen lokaci-lokaci mataki ne mai wayo, mai amfani don zuwa ga amincin ƙira, babban cikas ga amfani da masana'antu.
  • Hanyar Mai Da Hankali kan Fasali: Ƙarfafa kan fasalin tattalin arziki na asali (ciniki, farashin musayar kuɗi) akan nazarin fasaha kawai yana kafa ƙirar a cikin gaskiyar tattalin arziki.
  • Ƙwaƙƙwaran Kwatankwacin Aiki: Kwatanta LSTM, CNN, da Transformer yana ba da ma'auni na zamani mai amfani ga fagen.
Gazawar Muhimmi & Rashin Bayyanawa:
  • Haɗarin Wuce Gona da Irinsa An Yi Watsi Da Shi: Tare da fasali 40 da ƙirori masu rikitarwa, takarda mai yiwuwa ta fuskanci haɗari mai mahimmanci na wuce gona da iri. Cikakkun bayanai kan tsarin ƙa'ida (fita, lalacewar nauyi) da lokutan gwaji masu ƙarfi na samfurin (misali, ta hanyar rashin kwanciyar hankali na COVID-19) suna da mahimmanci kuma ba a ba da rahoto sosai ba.
  • Nuna Bambancin Bayanai: Tsarin zaɓin fasali, yayin da yake da tsauri, a zahiri yana gabatar da nuna bambanci idan ba a sarrafa shi da kyau tare da tagogin birgima ba. Wannan shine ƙafar Achilles na yawancin takardun kuɗi na ML.
  • Rashin Gwajin Girgizar Tattalin Arziki: Ta yaya TSMixer ya yi aiki a lokutan abubuwan da suka faru na baƙar fata? An lura da aikinsa a lokacin gyara na 2015, amma gwajin damuwa game da rushewar kasuwa na 2020 ko jujjuyawar Fed na 2022 zai fi bayyanawa.
  • Kwatanta da Ma'auni Masu Sauƙi: Shin ya fi ƙirar ARIMA mai sauƙi ko tafiya bazuwar aiki sosai? Wani lokaci, rikitarwa yana ƙara riba kaɗan a farashi mai yawa.

4.3 Fahimta Mai Amfani

Ga masu ƙididdiga da cibiyoyin kuɗi:

  1. Ba da Fifiko ga TSMixer don Ayyukan Gwaji: Daidaiton aiki da sauƙinsa sun sa ya zama farkon farko mai ƙarancin haɗari, babban riba don tsarin hasashen musayar kuɗi na cikin gida.
  2. Tilasta XAI don Tabbatar da Ƙira: Dage kan kayan aiki kamar Grad-CAM ba a matsayin tunani na baya ba, amma a matsayin babban ɓangare na tsarin ci gaban ƙira. Dole ne "hankali" na ƙira ya zama abin tantancewa kafin a tura shi.
  3. Mayar da Hankali kan Laburare na Fasali, Ba Kawai Ƙira Ba: Saka hannun jari a gina da kuma kiyaye ingantattun bayanai, masu ƙarancin jinkiri don rukunin fasali 6 da aka gano. Ƙirar tana da kyau kamar yadda man fetur ɗinta yake.
  4. Aiwatar da Ingantaccen Binciken Tsaka-tsakin Lokaci: Don yaƙi da nuna bambancin bayanai, ɗauki ƙa'idodin gwajin baya na asalin birgima kamar yadda aka bayyana a cikin binciken daga Bankin Tarayya (misali, aikinsu na yanzu).
Wannan takarda tsari ne, ba mafita mai sauƙi ba. Ƙimar sa ta gaske ita ce nuna hanyar da ke da ci gaba kuma mai ɗaukar alhaki.

5. Zurfin Fasaha

5.1 Tsarin Lissafi

Babbar matsalar hasashe an tsara ta a matsayin hasashen farashin musayar kuɗi na lokaci na gaba $y_{t+1}$ idan aka ba da jerin lokaci masu yawa na fasali $\mathbf{X}_t = \{x^1_t, x^2_t, ..., x^F_t\}$ a kan taga dubawa na $L$ lokaci: $\{\mathbf{X}_{t-L}, ..., \mathbf{X}_t\}$.

Layer na TSMixer (An Sauƙaƙe): Muhimmin aiki a cikin TSMixer ya haɗa da nau'ikan haɗuwa na MLP guda biyu:

  • Haɗuwa na Lokaci: $\mathbf{Z} = \sigma(\mathbf{W}_t \cdot \mathbf{X} + \mathbf{b}_t)$ yana amfani da sassa mai kauri a cikin girma na lokaci don kowane fasali da kansa, yana ɗaukar alamu na lokaci.
  • Haɗuwa na Fasali: $\mathbf{Y} = \sigma(\mathbf{W}_f \cdot \mathbf{Z}^T + \mathbf{b}_f)$ yana amfani da sassa mai kauri a cikin girma na fasali a kowane matakin lokaci, yana ƙirar hulɗa tsakanin alamomin tattalin arziki daban-daban.
inda $\sigma$ shine kunnawa mara tsari (misali, GELU), $\mathbf{W}$ sune matrices na nauyi, kuma $\mathbf{b}$ sune bambance-bambance.

Grad-CAM don Jerin Lokaci: Don hasashen manufa $\hat{y}$, maki mahimmanci $\alpha^c_k$ don fasali $k$ ana ƙididdige shi ta hanyar gradient backpropagation: $$\alpha^c_k = \frac{1}{T} \sum_{t} \frac{\partial \hat{y}^c}{\partial A^k_t}$$ inda $A^k_t$ shine kunnawar ƙarshe ko sassa mai kauri don fasali $k$ a lokacin $t$. Taswirar zafi na ƙarshe na Grad-CAM $L^c_{Grad-CAM}$ haɗuwa ce mai nauyin waɗannan kunnawa: $L^c_{Grad-CAM} = ReLU(\sum_k \alpha^c_k A^k)$. ReLU yana tabbatar da cewa kawai fasali masu tasiri mai kyau ne ake nuna.

5.2 Misalin Tsarin Nazari

Hali: Nazarin Hankalin Ƙira Yayin Sanarwar Manufa
Yanayi: Fed ya sanar da haɓaka ƙimar riba da ba zato ba tsammani. Ƙirar TSMixer ɗin ku tana hasashen raguwar darajar RMB.

  1. Mataki 1 - Samar da Hasashe & Grad-CAM: Kunna ƙirar don lokacin da ke biye da sanarwar. Cire taswirar zafi na Grad-CAM.
  2. Mataki 2 - Fassara Taswirar Zafi: Gano waɗanne layukan fasali (misali, `USD_Index`, `CN_US_Interest_Diff`) suka nuna kunnawa mai yawa (ja) a lokacin sanarwar da nan da nan bayan matakin lokacin sanarwar.
  3. Mataki 3 - Tabbatarwa da Hankali: Shin hankalin ƙirar ya yi daidai da ka'ida? Ƙarfafa hankali kan bambance-bambancen ƙimar riba yana tabbatar da ƙirar. Idan ya fi mayar da hankali kan, a ce, `Farashin_Mai`, zai ɗaga tuta ja da ke buƙatar bincike kan haɗin kai mara kyau.
  4. Mataki 4 - Aiki: Idan an tabbatar, fahimtar tana ƙarfafa amincewa da amfani da ƙirar don nazarin yanayi game da tarurrukan Fed na gaba. Taswirar zafi yana ba da rahoto kai tsaye, na gani ga masu ruwa da tsaki.
Wannan tsarin yana juya tambayar ƙira daga aikin ƙididdiga zuwa tsari mai tsari, mai fahimta.

6. Ayyuka na Gaba & Jagorori

Hanyar da aka fara a nan tana da fa'ida mai faɗi fiye da RMB/USD:

  • Hasashen Kadara Da Yawa: Yin amfani da TSMixer+Grad-CAM ga wasu nau'ikan musayar kuɗi, rashin kwanciyar hankali na cryptocurrency, ko hasashen farashin kayayyaki.
  • Nazarin Tasirin Manufa: Bankunan tsakiya za su iya amfani da irin waɗannan ƙirori masu bayyanawa don kwaikwayi tasirin kasuwa na yiwuwar canje-canjen manufofi, fahimtar waɗanne tashoshi (ƙimar riba, jagorar gaba) kasuwa ta fi kula da su.
  • Gudanar da Haɗari na Ainihi: Haɗa wannan bututun a cikin allunan kasuwanci na ainihi, inda Grad-CAM ke haskaka canji a cikin abubuwan da ke motsa su yayin da labarai ke watse, yana ba da damar daidaita dabarun kariya mai motsi.
  • Haɗawa tare da Bayanai Madadin: Aikin nan gaba dole ne ya haɗa da bayanai marasa tsari (halin labarai daga ƙirar NLP, sautin magana na bankin tsakiya) a matsayin ƙarin fasali, ta amfani da tsarin bayyanawa iri ɗaya don auna tasirinsu akan tushen gargajiya.
  • Gano Dalili: Gaba gaba shine motsawa daga haɗin kai (wanda Grad-CAM ya haskaka) zuwa dalili. Fasaha kamar algorithms na gano dalili (misali, PCMCI) za a iya haɗa su tare da ƙirar DL don bambance masu motsa asali daga alamu na haɗin kai.

7. Nassoshi

  1. Meng, S., Chen, A., Wang, C., Zheng, M., Wu, F., Chen, X., Ni, H., & Li, P. (2023). Haɓaka Hasashen Canjin Kuɗi tare da Ƙirar Koyon Zurfi Masu Bayyanawa. Rubutun da ake shirya.
  2. Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Bayyanawa ta Hotuna daga Cibiyoyin Sadarwa masu zurfi ta hanyar Gano Wuri ta hanyar Gradient. Proceedings of the IEEE International Conference on Computer Vision (ICCV), 618-626.
  3. Chen, S., & Hardle, W. K. (2023). AI a cikin Kuɗi: Kalubale, Ci gaba, da Damammaki. Annual Review of Financial Economics, 15.
  4. Bankin Tarayya na New York. (2022). Yin Hasashe Yanzu tare da Manyan Bayanai. Rahoton Ma'aikata. An samo daga https://www.newyorkfed.org/research/staff_reports
  5. Diebold, F. X., & Yilmaz, K. (2015). Haɗin Kuɗi da Tattalin Arziki: Hanyar Sadarwa don Aunawa da Sa ido. Oxford University Press.