3.1. Shirya Bayanai
Ana tsaftace bayanan Forex na danye, a daidaita su, kuma a tsara su zuwa matakan lokaci masu bi da bi wadanda suka dace da shigarwar LSTM. Injiniyan fasali na iya haɗawa da alamomin fasaha (misali, matsakaicin motsi, RSI).
1. Gabatarwa
Kasuwancin Musanya Kudi na Waje (Forex), wanda ke da girman ciniki na yau da kullun wanda ya wuce dala tiriliyan 5, yana wakiltar kasuwar kuɗi mafi girma kuma mafi ruwa a duniya. Yin hasashen daidai na farashin musanya kudi, musamman ga manyan nau'ikan kamar EUR/USD, yana da mahimmanci ga sarrafa haɗari da haɓaka riba. Wannan binciken yana binciken aikace-aikacen hanyoyin sadarwar jijiyoyi na Dogon Ƙwaƙwalwar Ƙwaƙwalwa (LSTM) don wannan aikin, tare da mai da hankali biyu: inganta daidaiton hasashe da kimanta tasirin tsarin akan amfani da makamashin lissafi. Binciken yana nufin haɗa hasashen kuɗi tare da ayyukan lissafi masu dorewa.
2. Bita na Adabi
Hasashen Forex ya samo asali daga tsoffin hanyoyin bincike na fasaha da na asali zuwa ingantattun dabarun koyon inji. Tsofaffin tsare-tsare sun dogara da hanyoyin jeri na lokaci na ƙididdiga (misali, ARIMA). Zuwan hanyoyin sadarwar jijiyoyi na wucin gadi (ANNs) da injunan tallafi vector (SVMs) sun nuna wani sauyi mai mahimmanci. Kwanan nan, tsare-tsaren koyo mai zurfi, musamman LSTM da haɗaɗɗun su (misali, LSTM-RCN), sun sami shahara saboda ikonsu na ɗaukar dogon lokaci na dogaro a cikin bayanan kuɗi masu sauyi—wata fa'ida mai mahimmanci fiye da ƙananan tsare-tsare.
3. Hanyoyi & Tsarin Tsari
Binciken yana amfani da hanyar koyo mai kulawa ta amfani da bayanan tarihin farashin musanya EUR/USD.
3.2. Zane na Tsarin LSTM
An tsara tsarin LSTM mai yawan sassa. Tsarin ya haɗa da sassan LSTM don sarrafa jerin gwano, sannan kuma sassan Dense don hasashen fitarwa. Ana daidaita ma'auni kamar adadin sassa, raka'a, da ƙimar zubewa.
3.3. Ma'auni na Kimantawa
Ana tantance aikin tsarin sosai ta amfani da ma'auni guda uku masu mahimmanci:
- Matsakaicin Kuskuren Square (MSE): $MSE = \frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2$
- Matsakaicin Kuskure na Cikakke (MAE): $MAE = \frac{1}{n}\sum_{i=1}^{n}|y_i - \hat{y}_i|$
- R-squared (R²): $R^2 = 1 - \frac{\sum_{i}(y_i - \hat{y}_i)^2}{\sum_{i}(y_i - \bar{y})^2}$
4. Sakamakon Gwaji & Bincike
4.1. Ma'auni na Aiki
Tsarin LSTM da aka inganta, wanda aka horar da shi na tsawon zamani 90, ya nuna aiki mafi girma idan aka kwatanta da tsare-tsaren tushe (misali, RNN mai sauƙi, ARIMA). Sakamako masu mahimmanci sun haɗa da:
- Ƙananan ƙimar MSE da MAE, suna nuna babban daidaiton hasashe ga motsin farashin EUR/USD.
- Ƙimar R² kusa da 1, tana nuna cewa tsarin ya bayyana babban yanki na bambance-bambancen a cikin bayanan farashin musanya.
- Tsarin ya ɗauki ingantattun alamu, marasa layi da kuma dogon lokaci na yanayin kasuwar Forex.
4.2. Binciken Amfani da Makamashi
Binciken ya nuna wani muhimmin al'amari, wanda sau da yawa ake yin watsi da shi: farashin lissafi na koyo mai zurfi. Horar da tsare-tsaren LSTM masu rikitarwa yana buƙatar albarkatun GPU/CPU masu mahimmanci, wanda ke haifar da yawan amfani da makamashi. Takardar tana jayayya cewa inganta tsarin (misali, ingantaccen tsari, tsayawa da wuri a zamani 90) ba wai kawai yana inganta daidaito ba har ma yana rage nauyin lissafi, don haka yana rage alamar makamashi da ke tattare da shi, kuma yana ba da gudummawa ga dorewar muhalli a cikin cinikin algorithm.
5. Babban Fahimta & Ra'ayi na Manazarta
Babban Fahimta: Ainihin ƙimar wannan takarda ba wai kawai wani sakamako na "LSTM ya doke tushe a cikin kuɗi" ba ne. Babban fahimtarta ita ce tsara inganta tsarin a matsayin matsalar manufa biyu: haɓaka ƙarfin hasashe yayin rage yawan kuɗin lissafi na makamashi. A cikin zamanin da sawun carbon na AI ke ƙarƙashin bincike (kamar yadda aka nuna a cikin bincike kamar waɗanda suka fito daga aikin ML CO2 Impact), wannan yana canza manufar daga kawai daidaito zuwa daidaito mai inganci.
Kwararren Tsari: Muhawarar tana ci gaba da ma'ana: 1) Hasashen Forex yana da ƙima amma yana da ƙarfi a lissafi. 2) LSTM sune mafi kyawun zamani don hasashen jerin gwano. 3) Za mu iya inganta su (tsari, zamani). 4) Ingantawa yana inganta ma'auni (MSE, MAE, R²). 5) Mafi mahimmanci, wannan ingantawar guda ɗaya tana rage maimaita lissafi, tana adana makamashi. 6) Wannan ya yi daidai da ƙarin ka'idojin AI na Kore. An yi haɗin kai tsakanin ingancin tsari da ingancin makamashi da gamsarwa.
Ƙarfi & Kurakurai: Ƙarfi: Kusurwar tsaka-tsakin ilimi yana da hankali kuma yana da buƙatu. Yana haɗa fasahar kuɗi tare da lissafi mai dorewa. Amfani da ma'auni na yau da kullun (MSE, MAE, R²) yana sa da'awar aikin ta zama mai tabbatarwa. Kuskure Mai Mahimmanci: Takardar tana da haske sosai akan ƙididdige adadin makamashin da aka adana. Ta ambaci ra'ayi amma ba ta da bayanai masu ƙarfi—babu joules da aka adana, babu daidaitaccen carbon da aka rage, babu kwatanta amfani da makamashi a kowane zamani. Wannan babbar dama ce da aka rasa. Ba tare da wannan ƙididdiga ba, hujjar makamashi ta kasance ta inganci kuma mai ba da shawara maimakon ƙarshe. Bugu da ƙari, ƙarfin tsarin ga matsanancin abubuwan kasuwa ("baƙar fata") ba a magance shi ba—wani gibi mai mahimmanci ga tsarin ciniki na ainihin duniya.
Fahimta Mai Aiki: Ga masu ƙididdiga da ƙungiyoyin AI: 1) Kula da Horonku: Nan da nan fara bin jan ƙarfin GPU (ta amfani da kayan aiki kamar NVIDIA-SMI) tare da ma'auni na asara. Kafa ma'auni na "aiki a kowace watt". 2) Wuce Tsayawa da wuri: Gwada ƙarin dabarun inganci kamar tsinke tsarin, ƙididdigewa (kamar yadda aka bincika a cikin TensorFlow Lite), ko distillation ilimi don ƙirƙirar ƙananan tsare-tsare, masu sauri, ƙarancin makamashi waɗanda ke riƙe da daidaito. 3) Gwada Ƙarfi don Ƙarfi: Tabbatar da tsarin ba kawai a lokutan al'ada ba amma akan bayanan rikici masu yawan canji. Tsarin da ya gaza a shiru yayin rushewar kasuwa ya fi mara amfani. Nan gaba na tsare-tsare ne waɗanda suke da wayo kuma masu inganci.
6. Cikakkun Bayanai na Fasaha & Tsarin Lissafi
Tsakiyar tantanin LSTM tana magance matsalar gradient da ke ɓacewa ta hanyar tsarin ƙofar. Maɓallin lissafi don lokaci guda (t) sune:
Ƙofar Manta: $f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$
Ƙofar Shigarwa: $i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$
Matsayin Tantamin ɗan Takara: $\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)$
Sake Sabunta Matsayin Tanti: $C_t = f_t * C_{t-1} + i_t * \tilde{C}_t$
Ƙofar Fitarwa: $o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$
Fitar Matsayin Boye: $h_t = o_t * \tanh(C_t)$
Inda $\sigma$ shine aikin sigmoid, $*$ yana nuna ninka kashi-kashi, $W$ da $b$ sune ma'auni da son kai, $h$ shine matsayin ɓoye, kuma $x$ shine shigarwa.
Aikin asarar tsarin yayin horo yawanci shine Matsakaicin Kuskuren Square (MSE), kamar yadda aka bayyana a baya, wanda mai ingantawa (misali, Adam) yana ragewa ta hanyar daidaita ma'auni (W, b).
7. Tsarin Bincike: Wani Lamari na Aiki
Yanayi: Asusun shinge na ƙididdiga yana son haɓaka siginar ciniki mai ƙarancin jinkiri, mai hankali ga makamashi don EUR/USD.
Aikace-aikacen Tsarin:
- Ma'anar Matsala: Yi hasashen shugaban fitila na awanni 4 na gaba (sama/ƙasa) tare da daidaito >55%, tare da lokacin shigar tsarin < 10ms da manufar rage makamashin horo da 20% idan aka kwatanta da LSTM na tushe.
- Bayanai & Shirya: Yi amfani da bayanan OHLCV na awa 5 na shekaru. Ƙirƙiri fasali: dawowar log, tagogin jujjuyawar jujjuyawar, da wakilai na rashin daidaiton littafin oda. Daidaita kuma a jera su cikin tagogin matakan lokaci 50.
- Zane na Tsari Mai Inganci: Fara da ƙaramin LSTM (misali, raka'a 32). Yi amfani da Optimization na Bayesian don daidaita ma'auni (sassa, zubewa, ƙimar koyo) tare da haɗaɗɗun aikin manufa: (Daidaito * 0.7) + (1 / Amfani da Makamashi * 0.3). Aiwatar da tsayawa da wuri tare da haƙuri na zamani 15.
- Kimantawa & Aiwalewa: Kimanta akan saitin gwaji da aka ajiye don daidaito, ma'auni na Sharpe na dabarar kwaikwayo, da auna lokacin shigarwa/iko. Tsarin ƙarshe shine sigar da aka tsinke na LSTM mafi kyau, wanda aka aiwatar ta hanyar Sabis na TensorFlow don aiwatarwa mai inganci.
8. Aikace-aikace na Gaba & Hanyoyin Bincike
- AI na Kore don Kuɗi: Haɓaka ma'auni na yau da kullun don "Ingancin Makamashi a kowace Rago na Hasashe" a cikin tsare-tsaren kuɗi. Matsawa na tsari don bayyana sawun carbon na AI a cikin rahotannin ESG.
- Haɗaɗɗun & Tsare-tsare masu Sauƙi: Bincike cikin haɗa LSTM tare da hanyoyin hankali (Masu Canzawa) don mafi kyawun mai da hankali na dogon zango, ko amfani da ingantattun tsare-tsare kamar Cibiyoyin Sadarwa na Lokaci (TCNs) ko Cibiyoyin Sadarwa na Lokaci na Ruwa (LTCs) don yuwuwar ƙarancin farashin lissafi.
- AI Mai Bayyanawa (XAI): Haɗa dabarun kamar SHAP ko LIME don bayyana hasashen LSTM Forex, gina amincewar ɗan ciniki da cika yuwuwar buƙatun tsari don bayyanawa.
- Shigar da Tsakiya & Bakin Gefe: Aiwalda ingantattun tsare-tsare don hasashe akan na'urori na gefe kusa da sabobin ciniki, rage jinkirin canja wurin bayanai da makamashi.
- Hasashen Kadari da Kasuwa: Faɗaɗa tsarin don hasashen alaƙa tsakanin EUR/USD da sauran nau'ikan kadari (misali, fihirisar adalci, kayayyaki) don sarrafa haɗarin matakin fayil.
9. Nassoshi
- Hochreiter, S., & Schmidhuber, J. (1997). Dogon Ƙwaƙwalwar Ƙwaƙwalwa. Lissafi na Jijiyoyi, 9(8), 1735–1780.
- Sejnowski, T. J., et al. (2020). Sawun Carbon na AI da Koyon Injin. Sadarwar ACM.
- Bankin don Hukumomin Tsakiya na Duniya (BIS). (2019). Binciken Bankin Tsakiya na Shekaru Uku na Musanya Kudi na Waje da Kasuwannin OTC Derivatives.
- Zhu, J.-Y., et al. (2017). Fassarar Hotuna zuwa Hotuna marasa Haɗin gwiwa ta amfani da Cibiyoyin Adawa na Haɗin kai. Taron Ƙasa da Ƙasa na IEEE akan Kwamfuta na Gani (ICCV). (CycleGAN a matsayin misali na sabon tsarin koyo mai zurfi).
- Strubell, E., Ganesh, A., & McCallum, A. (2019). Makamashi da La'akari da Manufofin don Koyo Mai Zurfi a cikin NLP. Gudanar da Taron Shekara-shekara na 57 na Ƙungiyar Lissafi na Kwamfuta.
- Kayan Aikin Inganta Tsarin TensorFlow. (n.d.). An samo daga https://www.tensorflow.org/model_optimization