Haɓaka Hasashen Farashin Musayar Dala da Taka na Bangladesh (USD/BDT) ta Amfani da LSTM da Koyon Injin

1. Gabatarwa

Hasashen da ya dace na farashin musayar Dala na Amurka zuwa Taka na Bangladesh (USD/BDT) yana da mahimmanci ga tattalin arzikin Bangladesh wanda ya dogara da shigo da kayayyaki, yana shafar ma'auni na ciniki, hauhawar farashin kayayyaki, da sarrafa ajiyar kuɗin waje. Tsarin ƙididdiga na gargajiya sau da yawa sun kasa kama tsarin da ba na layi ba, rikitattun tsarin da ke da alaƙa da kuɗaɗen kasuwannin masu tasowa, musamman a ƙarƙashin rashin tabbas na tattalin arziki. Wannan binciken yana magance wannan gibi ta hanyar haɓakawa da kimanta ƙwararrun samfuran koyon injin, musamman cibiyoyin sadarwa na Ƙwaƙwalwar Lokaci Mai Tsayi (LSTM) da Masu Rarraba Haɓakar Gradient (GBC), ta amfani da bayanan tarihi daga 2018 zuwa 2023. Binciken yana da nufin samar da ingantattun kayan aiki don rage haɗarin kuɗi da tsara manufofi.

2. Bita na Adabi

Aikace-aikacen koyo mai zurfi, musamman cibiyoyin sadarwa na LSTM, sun nuna alƙawari mai mahimmanci a cikin hasashen jerin lokaci na kuɗi. Hochreiter & Schmidhuber sun fara shi don warware matsalar gradient da ke ɓacewa a cikin RNNs, LSTMs sun yi fice wajen kama dogon lokaci masu dogaro. Haɓakawa na gaba kamar ƙofofin mantawa (Gers et al.) sun inganta daidaitawa ga sauyi. Nazarin gwaji, kamar waɗanda ke kan USD/INR, sun nuna LSTMs sun fi tsarin ARIMA na gargajiya da 18-22% cikin daidaiton shugabanci. Duk da haka, binciken da ke mai da hankali musamman kan nau'in USD/BDT, la'akari da tsarin gudanar da tsarin float na musamman na Bangladesh da kuma girgizar tattalin arzikin gida, ya kasance mai iyaka. Wannan binciken ya ginu kuma ya faɗaɗa wannan fagen da ba a taɓa gani ba.

3. Hanyoyi & Bayanai

3.1 Tattara Bayanai & Shirye-shiryen Farko

Bayanan farashin musayar USD/BDT na yau da kullun daga Janairu 2018 zuwa Disamba 2023 an samo su daga Yahoo Finance. An tsaftace bayanan, kuma an ƙirƙira fasali kamar daidaitattun dawowar yau da kullun, matsakaicin motsi mai sauƙi (SMA), da fihirisar ƙarfin dangi (RSI) don kama yanayin kasuwa da sauyi. An raba bayanan zuwa saiti na horo (80%) da gwaji (20%).

3.2 Tsarin Tsarin LSTM

Babban samfurin hasashen shine cibiyar sadarwa ta LSTM da aka tara. Tsarin yawanci ya ƙunshi:

• Layer na Shigarwa: Jerin bayanan farashi/fasali na tarihi.
• Layers na LSTM: Layer biyu ko fiye tare da jujjuyawa don daidaitawa don hana wuce gona da iri.
• Layer Mai Yawa: Layer mai cikakken haɗin kai don fitarwa.
• Layer na Fitarwa: Neuron guda ɗaya don hasashen farashin musayar lokaci na gaba.

An horar da samfurin ta amfani da mai inganta Adam da Matsakaicin Kuskuren Murabba'i (MSE) azaman aikin asara.

3.3 Na'urar Rarraba Haɓakar Gradient

Don hasashen shugabanci (motsi sama/ƙasa), an aiwatar da Mai Rarraba Haɓakar Gradient (GBC). Yana amfani da tarin raunannun samfuran hasashe (bishiyoyin yanke shawara) don ƙirƙirar mai rarraba mai ƙarfi, yana mai da hankali kan rage kuskuren hasashe ta hanyar koyo mai maimaitawa.

4. Sakamakon Gwaji & Nazari

4.1 Ma'aunin Aiki

Samfurin LSTM ya sami sakamako na musamman: daidaito na 99.449%, Tushen Matsakaicin Kuskuren Murabba'i (RMSE) na 0.9858, da asarar gwaji na 0.8523. Wannan aikin ya fi na tsarin ARIMA na gargajiya, wanda yake da RMSE na 1.342. Babban daidaito yana nuna babban iyawar LSTM a cikin ƙirar rikitattun yanayin lokaci na farashin musayar USD/BDT.

4.2 Gwajin Baya & Kwaikwayon Ciniki

An gwada Mai Rarraba Haɓakar Gradient a kan kwaikwayon ciniki wanda ya fara da babban jari na $10,000. A kan ciniki 49, samfurin ya sami yawan ciniki mai riba na 40.82%. Duk da haka, kwaikwayon ya haifar da asarar net na $20,653.25. Wannan yana nuna hasashe mai mahimmanci: babban daidaiton shugabanci ba ya canzawa kai tsaye zuwa dabarun ciniki masu riba, kamar yadda farashin ma'amala, zamewa, da sarrafa haɗari (matakan dakatar da asara/ɗaukar riba ba a ambata a cikin PDF ba) suna taka muhimmiyar rawa.

Bayanin Chati (An fahimta): Chati na layi zai iya nuna tarihin ƙimar USD/BDT yana raguwa daga kusan 0.012 (2018) zuwa 0.009 (2023). Chati na biyu zai zana tarin P&L na dabarun ciniki na GBC, yana nuna lokacin farko na riba wanda ya biyo bayan ja da baya mai zurfi wanda ya kai ga asarar net ta ƙarshe.

5. Zurfin Fasaha

Tushen ingancin LSTM yana ta'allaka ne a cikin yanayin tantanin halitta da hanyoyin ƙofa. Maɓallin lissafi don tantanin LSTM a lokacin mataki $t$ sune:

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

Inda $\sigma$ shine aikin sigmoid, $*$ yana nuna ninka abubuwan abu, $W$ da $b$ sune ma'auni da son kai, $x_t$ shine shigarwa, $h_t$ shine yanayin ɓoye, kuma $C_t$ shine yanayin tantanin halitta. Wannan tsarin yana ba da damar samfurin don tunawa ko mantawa da bayanai a kan dogon jerin gwano, mai mahimmanci ga jerin lokaci na kuɗi tare da dogon lokaci masu dogaro.

6. Tsarin Nazari & Misalin Lamari

Tsarin: Bututun ML na Forex
Wannan binciken yana misalta daidaitaccen bututun mai tasiri don ML na kuɗi:

1. Tsara Matsala: Koma baya (LSTM don farashi) vs. Rarrabuwa (GBC don shugabanci).
2. Injiniyan Fasali: Ƙirƙirar siginonin hasashe daga farashin danye (dawowa, alamomin fasaha).
3. Zaɓin Samfuri & Horarwa: Zaɓar samfuran masu sanin jerin gwano (LSTM) don bayanan lokaci.
4. Ingantaccen Tabbatarwa: Amfani da ƙetare-lokaci, ba raba bazuwar ba, don guje wa son zuciya mai kallon gaba.
5. Dabarun Gwajin Baya: Fassara hasashen samfurin zuwa dabarun ciniki na kwaikwayo tare da ƙayyadaddun gaske.

Misalin Lamari: Samar da Siginoni
Ƙa'idar da aka sauƙaƙa bisa hasashen LSTM na iya zama: "Idan farashin da aka hasashe don gobe ya fi (farashin yau + bakin kofa $\alpha$), samar da siginonin SAYE." GBC kai tsaye yana fitar da alamar aji (1 don SAMA, 0 don ƘASA). Muhimmin darasi daga asarar cinikin takardar shine wajabcin wani Layer na sarrafa haɗari na gaba wanda ke ƙayyade girman matsayi, umarnin dakatar da asara, da rabon fayil, wanda wataƙila ba ya nan ko kuma mai sauƙi a cikin kwaikwayon.

7. Aikace-aikace na Gaba & Hanyoyi

Makomar AI a cikin hasashen forex yana ta'allaka ne a cikin tsarin daidaitawa, masu daidaitawa:

• Haɗaɗɗun Bayanan Madadin: Haɗa nazarin yanayin labarai na ainihi (ta amfani da samfuran NLP kamar BERT), sautin sadarwar babban banki, da fihirisar haɗarin siyasa, kamar yadda aka gani a cikin asusun shinge kamar Two Sigma.
• Samfuran Haɗin kai & Masu Hankali: Matsawa bayan daidaitattun LSTMs zuwa tsarin Transformer tare da hanyoyin kula da kai (kamar waɗanda ke cikin Vaswani et al.'s "Hankali shine Duk Abin da Kuke Bukata") waɗanda zasu iya auna mahimmancin matakan lokaci daban-daban da sassauci.
• Koyo mai Ƙarfafawa (RL): Haɗaɗɗun wakilan RL waɗanda ke koyon ingantattun manufofin ciniki kai tsaye, la'akari da farashi da dawowar da aka daidaita haɗari, maimakon kawai hasashen farashi. Wannan ya yi daidai da bincike daga DeepMind da OpenAI a cikin yanayin kwaikwayo.
• AI Mai Bayyanawa (XAI): Aiwar fasahohi kamar SHAP ko LIME don fassara hasashen samfurin, wanda ke da mahimmanci don bin ka'idoji da samun amincewa daga cibiyoyin kuɗi.
• Koyo na Kasuwa-Kasuwa: Horar da samfura akan nau'ikan kuɗi da yawa ko nau'ikan kadari don koyon tsarin gama gari na sauyi da yaɗuwa.

8. Nassoshi

1. Hochreiter, S., & Schmidhuber, J. (1997). Ƙwaƙwalwar Lokaci Mai Tsayi. Lissafin Neuronal.
2. Gers, F. A., Schmidhuber, J., & Cummins, F. (2000). Koyon Mantawa: Hasashe na Ci gaba tare da LSTM.
3. Rahman et al. (2022). Hasashen Tushen LSTM don Kuɗaɗen Kasuwa Masu Tasowa: Nazarin Lamarin USD/INR. Jaridar Lissafin Kuɗi.
4. Afrin, S., et al. (2021). Hasashen Farashin Musayar USD/BDT ta Amfani da Koyon Injin. Taron Ƙasa da Ƙasa akan Kwamfuta da Fasahar Bayanai.
5. Vaswani, A., et al. (2017). Hankali Shine Duk Abin da Kuke Bukata. Ci gaba a cikin Tsarin Bayanai na Neuronal.
6. Yahoo Finance. (2023). Bayanan Tarihi na USD/BDT.

9. Ra'ayin Mai Nazarin Masana'antu

Babban Hasashe: Wannan takarda misali ne na al'ada na "sabani na daidaito-riba" a cikin kuɗaɗen ƙididdiga. Marubutan sun gina ingantaccen samfurin LSTM wanda ya cimma daidaito kusan cikakke na 99.45% akan hasashen USD/BDT—wani abin yabo—duk da haka dabarun ciniki da suka haɗa sun zubar da jari cikin bala'i. Labarin gaske ba shine daidaiton samfurin ba; shine babban rabuwa tsakanin ingantaccen ma'auni na ilimi da ciniki na gaske P&L. Ya jaddada gaskiyar da yawa masu ƙididdiga suka koya da wahala: rage RMSE ba daidai yake da haɓaka Ratio Sharpe ba.

Kwararar Ma'ana: Binciken yana bin daidaitaccen bututun: samun bayanai, injiniyan fasali, zaɓin samfuri (LSTM/GBC), da tabbatar da aiki. Kuskuren ma'ana, duk da haka, yana cikin tsalle daga tabbatarwa zuwa aikace-aikace. Gwajin baya ya bayyana maras hankali, wataƙila ba shi da ingantaccen samfurin farashin ma'amala, zamewa, kuma mafi mahimmanci, ingantaccen tsarin sarrafa haɗari. Yawan nasara na 40% tare da babban sakamako mara kyau yana nuna cewa asarar dabarar a kowane ciniki da aka yi asara ya fi ribar kowane ciniki da aka yi nasara—wani aibi mara kyau wanda babu adadin daidaiton LSTM zai iya gyara.

Ƙarfi & Aibobi:

• Ƙarfi: Kyakkyawan injiniyan samfuri don nau'in kuɗi na musamman, maras bincike (USD/BDT). Kwatanta da ARIMA yana ba da ma'auni bayyananne. Bayyanannen ambaton asarar ciniki yana da gaskiya ta hankali kuma ya fi daraja fiye da yawancin takardu waɗanda kawai ke nuna nasarori.
• Aibobi: Kwaikwayon ciniki shine ainihin tunani na baya, yana bayyana rashin haɗin kai tsakanin hasashe da matakan aiwatarwa—zuciyar ciniki na tsari. Babu tattaunawa game da girman matsayi (misali, Ma'aunin Kelly), dakatar da asara, ko mahallin fayil. Bugu da ƙari, yayin da LSTMs suke da ƙarfi, yanayin akwatin baƙar fata har yanzu babban shamaki ne ga karɓa a cikin cibiyoyin kuɗi masu tsari idan aka kwatanta da ƙarin tarin fassara kamar Bishiyoyin Haɓakar Gradient.

Hasashe Mai Aiki:

1. Gina Gadar tare da Koyo mai Ƙarfafawa: Maimakon ɗaukar hasashe da ciniki a matsayin matakai daban-daban, aikin gaba ya kamata ya yi amfani da Koyo mai Ƙarfafawa (RL) daga ƙarshe zuwa ƙarshe. Wakilin RL, kama da waɗanda DeepMind ke amfani da su don wasan wasa, na iya koyon inganta don ma'aunin ciniki kai tsaye (misali, dawowar tarawa, rabo Sortino) daga bayanan danye, yana haɗa farashi da haɗari a cikin su.
2. Karɓi "Hasashe-Aiwatarwa-Haɗari" Trinity: Duk wani binciken hasashe dole ne a kimanta shi a cikin triad. Samfurin hasashe kawai shine kusurwa ɗaya. Dole ne a yi amfani da ƙarfi daidai ga samfurin aiwatarwa (tasirin kasuwa, farashi) da samfurin haɗari (VaR, ƙarancin da ake tsammani, sarrafa ja da baya).
3. Mayar da hankali kan Gano Tsarin Mulki: USD/BDT, a ƙarƙashin wani mai gudanar da iyo, yana da tsarin mulki daban-daban (kwanciyar hankali, shiga tsakani, rikici). Ya kamata a yi amfani da samfura kamar Samfuran Sauyin Markov ko algorithms na clustering don gano tsarin mulki na yanzu da farko, sannan a yi amfani da mafi dacewar samfurin hasashe. Hanyar samfurin ɗaya-daidai-duk yana da gajeriyar hangen nesa.
4. Ba da fifiko ga Bayyanawa: Don matsawa daga aikin ilimi zuwa kayan aikin ɗan ciniki, aiwatar da fasahohin XAI. Nuna wa ɗan ciniki cewa siginonin "sayarwa" 60% yana motsa shi ta hanyar faɗaɗa gibin ciniki kuma 40% ta hanyar rarrabuwar RSI yana gina amincewa fiye da akwatin baƙar fata mai daidaito 99%.