ﻃﺮاﺣﯽ ﻣﺪﻟﯽ ﺟﻬﺖ ﭘﯿﺶﺑﯿﻨﯽ ﺑﺎزده ﺷﺎﺧﺺ ﮐﻞ ﺑﻮرس اوراق ﺑﻬﺎدار)ﺑﺎ ﺗﺎﮐﯿﺪ ﺑﺮ ﻣﺪلﻫﺎي ﺗﺮﮐﯿﺒﯽ ﺷﺒﮑﻪ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ و ﻣﺪلﻫﺎي ﺧﺎﻧﻮاده GARCH(

Designing a Model for Forecasting the Stock Exchange Total Index Returns (Emphasizing on Combined Deep Learning Network Models and GARCH Family Models)

ﭼﮑﯿﺪه در ﺳﺎلﻫﺎي اﺧﯿﺮ، ﺗﻮﺳﻌﻪي ﭘﺮدازﻧﺪهﻫﺎي ﮐﺎﻣﭙﯿﻮﺗﺮي ﻣﻮﺟﺐ ﻣﻌﺮﻓﯽ اﻟﮕﻮرﯾﺘﻢﻫﺎي ﺟﺪﯾﺪي ﺑﺮاي ﭘﯿﺶﺑﯿﻨﯽ دادهﻫﺎي ﻣﺎﻟﯽ ﺷﺪه اﺳﺖ ﮐﻪ ﯾﮑﯽ از اﯾﻦ اﻟﮕﻮرﯾﺘﻢﻫﺎ، ﯾﺎدﮔﯿﺮي ﻣﺎﺷﯿﻦ )Machine Learning( اﺳﺖ. از اﯾﻨﺮو در ﭘﮋوﻫﺶ ﺣﺎﺿﺮ ﺑﻪ ﻣﻌﺮﻓﯽ ﯾﮏ ﻣﺪل ﺗﺮﮐﯿﺒﯽ از ﺷﺒﮑﻪ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ )Deep Learning( و ﻣﺪلﻫﺎي ﻣﻨﺘﺨﺐ ﺧﺎﻧﻮاده GARCH ﺟﻬﺖ ﭘﯿﺶﺑﯿﻨﯽ ﮐﻮﺗﺎهﻣﺪت ﺑﺎزدﻫﯽ روزاﻧﻪ ﺷﺎﺧﺺ ﮐﻞ ﺑﻮرس اوراق ﺑﻬﺎدار ﺗﻬﺮان ﭘﺮداﺧﺘﻪ ﻣﯽﺷﻮد. ﻣﻬﻤﺘﺮﯾﻦ وﯾﮋﮔﯽ ﺷﺒﮑﻪ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ در اﯾﻦ اﺳﺖ ﮐﻪ ﺑﺪون ﻣﺤﺪود ﺑﻮدن ﺑﻪ ﻣﺪلﻫﺎي ﻣﻌﯿﻦ، ﻣﯽﺗﻮاﻧﺪ ﺧﻮد را ﺑﺎ ﻧﻮﺳﺎﻧﺎت ﻣﺘﻐﯿﺮﻫﺎي ﺑﺎزار ﻫﻤﺎﻫﻨﮓ و ﺗﻌﺪﯾﻞ ﻧﻤﺎﯾﺪ. در اﯾﻦ ﭘﮋوﻫﺶ از ﻣﯿﺎن ﻣﺪلﻫﺎي ﺷﺒﮑﻪ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ، ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﺑﺎزﮔﺸﺘﯽ ﻣﺒﺘﻨﯽ ﺑﺮ ﺣﺎﻓﻈﻪ ﮐﻮﺗﺎهﻣﺪت و ﺑﻠﻨﺪﻣﺪت )RNN-LSTM( اﻧﺘﺨﺎب و از ﻣﺪلﻫﺎي داراي ﺣﺎﻓﻈﻪﮐﻮﺗﺎه ﻣﺪت GARCH و EGARCHدر ﺳﺎﺧﺘﺎر آن اﺳﺘﻔﺎده ﻣﯽﺷﻮد. ﻫﻤﭽﻨﯿﻦ دو ﻣﺘﻐﯿﺮ ﻣﺴﺘﻘﻞ ﻗﯿﻤﺖ ﻧﻔﺖ و ﻧﺮخ دﻻر در ﺳﺎﺧﺘﺎر ﻣﺪل ﺗﺮﮐﯿﺒﯽ، ﮐﻤﮏ ﻓﺮاواﻧﯽ ﺑﻪ آن در ﭘﯿﺶﺑﯿﻨﯽ دﻗﯿﻖﺗﺮ دادهﻫﺎي ﻣﺎﻟﯽ ﻣﯽﮐﻨﺪ. ﻧﺘﺎﯾﺞ ﺗﺤﻘﯿﻖ ﻧﺸﺎن ﻣﯽدﻫﺪ ﮐﻪ ﻣﺪلﻫﺎي ﺗﺮﮐﯿﺒﯽ دﻗﺖ ﭘﯿﺶﺑﯿﻨﯽ ﺑﺎﻻﺗﺮي ﻧﺴﺒﺖ ﺑﻪ ﻣﺪلﻫﺎي ﺗﮑﯽ دارﻧﺪ. ﻫﻤﭽﻨﯿﻦ ﺑﺮاﺳﺎس ﻣﻌﯿﺎرﻫﺎي ارزﯾﺎﺑﯽ ﺧﻄﺎي ﭘﯿﺶﺑﯿﻨﯽ RMSE و MAPE، ﻣﺪل RNN-LSTM-EGARCH ﺑﺮﭘﺎﯾﻪ ﺗﻮزﯾﻊ GED داراي ﺧﻄﺎي ﭘﯿﺶﺑﯿﻨﯽ ﮐﻤﺘﺮي ﻧﺴﺒﺖ ﺑﻪ 23 ﻣﺪل دﯾﮕﺮ دارد. در اﯾﻦ راﺳﺘﺎ، ﻣﻌﯿﺎر ﺑﺮرﺳﯽ ﺻﺤﺖ ﭘﯿﺶﺑﯿﻨﯽ دﯾﺒﻮﻟﺪ-ﻣﺎرﯾﺎﻧﻮ )DM( ﻧﯿﺰ ﯾﺎﻓﺘﻪﻫﺎي ﻓﻮق را ﺗﺎﯾﯿﺪ ﻣﯽﮐﻨﺪ.

Given the development of machine learning models in predicting financial data in recent years, this study introduces a combination of Deep Learning Network and selected GARCH family models to predict short-term daily returns of the Tehran Stock Exchange Index. The most important feature of the deep learning network is that it can adapt and adjust itself to the volatility of market variables without being limited to specific models. In this study, short-term and long-term memory based neural network (RNN-LSTM) models are used for deep learning network models and GARCH and EGARCH models are used in its structure. Also, the two independent variables of oil price and dollar rate in the structure of the hybrid model help to predict the financial data more accurately. Comparison of the results of hybrid model prediction error with individual models shows that the RNN-LSTM-EGARCH hybrid model has higher prediction accuracy than competing models. competing models.