Ci Gaba da Hasashen Canjin Kuɗi: Amfani da LSTM da AI don Hasashen USD/BDT

Teburin Abubuwan Ciki

1. Gabatarwa

Wannan binciken yana magance kalubalen da ke tattare da hasashen canjin kuɗin Dala na Amurka zuwa Taka na Bangladesh (USD/BDT), aiki mai mahimmanci ga tattalin arzikin Bangladesh mai dogaro da shigo da kayayyaki. Sauyin kuɗi yana shafar kai tsaye sarrafa ajiyar kuɗin waje, ma'auni na ciniki, da hauhawar farashin kayayyaki. Tsarin ƙididdiga na gargajiya sau da yawa ya kasa ɗaukar sifofin da ba su da layi, masu rikitarwa na kuɗaɗen kasuwannin masu tasowa, musamman a lokacin rashin tabbas na tattalin arziki. Wannan binciken yana amfani da ingantaccen koyon na'ura, musamman hanyoyin sadarwar jijiyoyi na LSTM, don ƙirƙirar waɗannan alaƙar lokaci mai ƙarfi ta amfani da bayanan tarihi daga 2018 zuwa 2023.

2. Bita na Adabi

Adabin kwanan nan ya tabbatar da fifikon hanyoyin sadarwar LSTM akan tsarin lokaci na gargajiya kamar ARIMA don hasashen kuɗi. Hochreiter & Schmidhuber sun ƙaddamar da su don warware matsalar gradient da ke ɓacewa a cikin RNNs, LSTMs sun yi fice wajen ɗaukar dogon lokaci na dogaro. Ƙarfafawa na gaba kamar ƙofofin mantawa (Gers et al.) sun inganta daidaitawa ga sauyin yanayi. Nazarin gwaji akan manyan nau'ikan kuɗi ya nuna LSTMs sun fi ARIMA da 18-22% cikin daidaiton shugabanci. Yayin da bincike akan kuɗaɗe kamar USD/INR ya wanzu, takamaiman nazarin kan USD/BDT yana da iyaka, sau da yawa ana amfani da bayanan kafin annobar cutar, kuma ba su da haɗakar fasahohin zamani kamar hanyoyin kulawa ko girgizar tattalin arzikin gida.

3. Hanyoyi & Bayanai

3.1. Tattara Bayanai & Shirye-shiryen Farko

An samo bayanan canjin kuɗin USD/BDT na yau da kullun na tarihi daga Yahoo Finance na tsawon lokacin 2018-2023. Bayanan sun nuna raguwar ƙimar BDT/USD daga kusan 0.012 zuwa 0.009. Shirye-shiryen bayanai sun haɗa da sarrafa ƙimar da ba a taɓa gani ba, ƙididdige daidaitattun ribar yau da kullun don ɗaukar sauyin yanayi, da ƙirƙirar jerin don tsarin lokaci.

3.2. Tsarin Tsarin LSTM

Babban tsarin hasashen shine hanyar sadarwar jijiyoyi ta LSTM. An inganta tsarin don bayanan USD/BDT, mai yiwuwa ya haɗa da yadudduka na LSTM da yawa, jujjuyawar don daidaitawa, da kuma ƙaƙƙarfan layin fitarwa. An horar da tsarin don hasashen ƙimar canjin kuɗi na gaba bisa jerin abubuwan da suka gabata.

3.3. Na'urar Rarraba Gradient Boosting (GBC)

An yi amfani da Na'urar Rarraba Gradient Boosting don hasashen shugabanci - hasashen ko ƙimar canjin kuɗi za ta tashi sama ko ƙasa. An kimanta aikin wannan tsarin ta hanyar kwaikwayon ciniki mai amfani.

4. Sakamakon Gwaji & Nazari

4.1. Ma'aunin Aikin LSTM

Tsarin LSTM ya sami sakamako na musamman: daidaito na 99.449%, Kuskuren Tushen Ma'anar Murabba'i (RMSE) na 0.9858, da asarar gwaji na 0.8523. Wannan yana nuna ingantaccen tsari don hasashen ainihin ƙimar ƙimar USD/BDT.

4.2. Kwaikwayon Ciniki na GBC

An gudanar da gwajin baya ta amfani da alamun shugabanci na GBC akan jarin farko na $10,000 akan cinikayya 49. Yayin da 40.82% na cinikayya sun kasance masu riba, dabarar ta haifar da asarar net na $20,653.25. Wannan yana nuna bambanci mai mahimmanci tsakanin daidaiton hasashe da ciniki mai riba, inda farashin ciniki, zamewa, da sarrafa haɗari suka fi mahimmanci.

4.3. Nazarin Kwatancen da ARIMA

Tsarin LSTM ya fi tsarin ARIMA na gargajiya gaba sosai, wanda yake da RMSE na 1.342. Wannan yana nuna fa'idar bayyananne na koyon zurfi a cikin ƙirƙirar sifofin da ba su da layi, masu rikitarwa da ke cikin bayanan lokaci na kuɗi.

5. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Kwayar LSTM tana aiki ta hanyar tsarin ƙofa wanda ke tsara kwararar bayanai. Maɓallin lissafin sune:

• Ƙofar Mantawa: $f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$
• Ƙofar Shigarwa: $i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$, $\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)$
• Sabunta Matsayin Kwaya: $C_t = f_t * C_{t-1} + i_t * \tilde{C}_t$
• Ƙofar Fitarwa: $o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$, $h_t = o_t * \tanh(C_t)$

Inda $\sigma$ shine aikin sigmoid, $*$ yana nuna ninkin kashi-kashi, $W$ sune matrices na nauyi, $b$ sune vectors na son kai, $x_t$ shine shigarwa, $h_t$ shine matsayin ɓoye, kuma $C_t$ shine matsayin kwaya. Wannan tsarin yana ba da damar hanyar sadarwa ta koyi wane bayani za ta riƙe ko watsar a cikin dogon jerin abubuwa.

6. Tsarin Nazari: Misali Mai Amfani

Harka: Haɗa Girgizar Tattalin Arzikin cikin Tsarin LSTM

Nazarin ya ambaci haɗa gano girgizar tattalin arzikin gida. Ga tsarin ra'ayi na yadda za a iya aiwatar da wannan ba tare da takamaiman lamba ba:

1. Ƙara Bayanai: Ƙirƙiri bayanan lokaci na "alamomin girgiza" na Bangladesh. Wannan zai iya zama tutocin binary (0/1) don abubuwan da suka faru kamar sanarwar shisshigin bankin tsakiya, manyan abubuwan siyasa, ko canje-canje a cikin kwararar kuɗin aiki, waɗanda aka samo daga APIs na labarai ko sanarwar hukuma.
2. Injiniyan Fasali: Ga kowace ranar ciniki, haɗa taga tarihin bayanan ƙimar canjin kuɗi tare da taga daidai na alamomin girgiza. Wannan yana haifar da ƙaƙƙarfan vector shigarwa: [Price_Seq, Shock_Seq].
3. Daidaitawar Tsari: Daidaita layin shigarwar LSTM don karɓar wannan shigarwa mai girma da yawa. Hanyar sadarwa za ta koyi haɗa takamaiman sifofin girgiza tare da sauyin yanayi ko canje-canjen yanayin a cikin ƙimar USD/BDT.
4. Tabbatarwa: Kwatanta aikin (RMSE, daidaiton shugabanci) na tsarin da aka ƙarfafa girgiza da tsarin tushe wanda ke amfani da bayanan farashi kawai, musamman a lokutan da aka yiwa alama da girgiza.

7. Aikace-aikace na Gaba & Hanyoyin Bincike

• Haɗin Bayanai Mai Yawa: Bayan tutocin tattalin arziki, haɗa nazarin ra'ayi na ainihin lokaci daga labaran kuɗi da kafofin watsa labarun zamantakewa (misali, ta amfani da tsarin Transformer kamar BERT) zai iya ɗaukar yanayin kasuwa, kamar yadda aka gani a cikin nazarin kan manyan nau'ikan kuɗi.
• Hanyoyin Kulawa: Haɗa yadudduka na kulawa (kamar waɗanda ke cikin tsarin Transformer) cikin LSTM zai iya ba da damar tsarin ya mai da hankali akai-akai akan mafi dacewa da matakan lokaci na baya, inganta fassarar da aiki don dogon jerin abubuwa.
• Koyon Ƙarfafawa don Ciniki: Matsawa daga hasashe mai tsafta zuwa koyon manufa kai tsaye. Tsarin kamar Deep Q-Network (DQN) za a iya horar da shi don yanke shawarar siye/sayar/riƙe waɗanda ke haɓaka ribar da aka daidaita haɗari (Sharpe Ratio), magance kai tsaye gibin riba da aka gani a cikin gwajin baya na GBC.
• Koyon Kuɗaɗen Ketare: Haɓaka tsarin meta da aka horar da shi akan nau'ikan kuɗaɗen kasuwannin masu tasowa da yawa (misali, USD/INR, USD/PKR) don koyon sifofin sauyin yanayi da tasirin manufa na duniya, sannan daidaitawa akan USD/BDT don ingantaccen ƙarfi tare da iyakance bayanai.

8. Nassoshi

1. Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation.
2. Gers, F. A., Schmidhuber, J., & Cummins, F. (2000). Learning to Forget: Continual Prediction with LSTM. Neural Computation.
3. Rahman et al. (Shekara). Nazarin kan hasashen USD/INR tare da LSTM. [Mujallar da ta dace].
4. Afrin et al. (2021). Nazarin kafin annobar cutar kan USD/BDT. [Taron da ya dace].
5. Hosain et al. (Shekara). Dabarun haɗin gwiwa don hasashen kuɗi. [Mujallar da ta dace].
6. Vaswani, A., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems.
7. Mnih, V., et al. (2015). Human-level control through deep reinforcement learning. Nature.

9. Asalin Nazari & Sharhin Kwararru

Babban Fahimta: Wannan takarda ta yi nasara wajen nuna fifikon fasaha na hanyoyin sadarwar LSTM akan tsarin da suka gabata kamar ARIMA don hasashen batu amma ba da gangan ba ta fallasa wani babban raguwa a cikin binciken fintech: haɗakar daidaiton ƙididdiga tare da amfanin tattalin arziki. Tsarin da ya kai 99.45% daidaito wanda, lokacin da aka fassara shi zuwa dabarar ciniki ta hanyar Na'urar Rarraba Gradient Boosting, ya haifar da asara mai ban tsoro na 200%+ akan jarin farko ba kawai bayanin kula na ilimi ba ne—yana kira ga canji na asali a yadda muke kimanta AI a cikin kuɗi.

Kwararar Hankali & Ƙarfafawa: Hankalin binciken yana da kyau kuma ana iya maimaitawa. Marubutan sun gano daidai iyakokin tsarin layi don kuɗaɗen da ba su da layi, masu kula da manufofi kamar BDT. Amfani da tsarin mulkin da aka sarrafa a matsayin nazarin shari'a yana da hikima, saboda waɗannan kasuwanni suna cike da rushewar AI. Aiwar fasaha tana da ƙarfi, tare da RMSE na LSTM na kusan cikakke na 0.9858 (vs. ARIMA's 1.342) yana ba da shaida maras gardama game da iyawar koyon zurfi don ƙirƙirar dogon lokaci na dogaro, binciken da ya yi daidai da ayyukan farko kamar takardar LSTM ta asali ta Hochreiter & Schmidhuber. Ƙoƙarin haɗawa zuwa sakamakon ciniki ta hanyar GBC mataki ne abin yabawa zuwa ga dacewar duniyar gaske.

Kurakurai Masu Mahimmanci & Sabani na Ribar: A nan ne babban aibi. Yawan nasara na GBC na 40.82% wanda ya haifar da asara mai yawa shine misali na al'ada na yin watsi da rashin daidaituwar ribar kuɗi. Yana nuna rashin haɗa ma'auni na haɗari (misali, Sharpe Ratio, Zazzagewa Mafi Girma) da kuma tsarin aiwatar da butulci. Wannan yayi daidai da rami na gama gari a cikin takardun AI na farko na kuɗi waɗanda suka mai da hankali kawai akan kuskuren hasashe. Fannin ya samo asali, kamar yadda aka gani a cikin hanyoyin ƙarfafawa waɗanda ke daidaita kai tsaye don ribar fayil, kamar tsarin Deep Q-Network (DQN) da aka yi amfani da shi a cikin babban aikin Mnih et al. Bugu da ƙari, yayin da takardar ta ambaci abubuwan tattalin arziki, aiwatar da ita da alama ba ta da kyau. Ga kuɗi kamar BDT, wanda shisshigin bankin tsakiya da kwararar kuɗin aiki ke tasiri sosai, rashin haɗa waɗannan a matsayin sifofi masu tsari—watakila ta amfani da tsarin kulawa don auna tasirin su, kamar yadda aka ba da shawara a cikin tsarin Transformer—damar da aka rasa.

Fahimta Mai Aiki & Hanyar Gaba: Ga masu aiki da masu bincike, wannan binciken yana ba da mahimman bayanai guda biyu masu aiki. Na farko, dakatar da bauta a bagadin RMSE. Babban ma'aunin kimantawa na kowane tsarin da ke fuskantar kasuwa dole ne ya zama aikin sa a cikin yanayin ciniki na kwaikwayo wanda ya haɗa da farashi na gaske, zamewa, da girman matsayi. Kayan aiki kamar Backtrader ko QuantConnect yakamata ba su zama abin sasantawa ba a cikin bututun tabbatarwa. Na biyu, gaba yana cikin koyon ƙarshe-zuwa-wakili. Maimakon bututun da ba a haɗa shi ba (LSTM -> GBC -> Ciniki), iyakar gaba ita ce amfani da wakili guda ɗaya, mai cikakken ra'ayi—mai yiwuwa ya dogara da Proximal Policy Optimization (PPO) ko makamantansu na RL—wanda ke shigar da bayanan kasuwa na danye ko an sarrafa su kaɗan kuma kai tsaye yana fitar da ayyukan ciniki da aka sarrafa haɗari. Aikin lada na wakilin zai zama haɗakar ma'auni na ribar da aka daidaita haɗari, tilasta wa AI ya koyi ainihin tattalin arzikin kasuwa, ba kawai sifofin ƙididdiga ba. Shawarar marubutan na ƙara nazarin ra'ayi fara mai kyau ne, amma dole ne a haɗa shi cikin wannan tsarin na tushen wakili, ba kawai a haɗa shi azaman ginshiƙin fasalin daban ba. Wannan shine hanyar daga ƙirƙirar mai hasashe mai wayo zuwa injiniyan wakili na kuɗi mai yuwuwa.