ارايه مدل بهينه ريسك اعتباري فرايند تامين مالي جمعي با استفاده از شبكه عصبي پرسپترون چندلايه (MLP)

Credit risk optimization model for crowdfunding process by using Neural Network(MLP

ﻫﺪف ﻣﻄﺎﻟﻌﻪ ﺣﺎﺿﺮ، ﭘﯿﺶ ﺑﯿﻨﯽ و اراﯾﻪ ﻣﺪل رﯾﺴﮏ اﻋﺘﺒﺎري ﺟﻬﺖ ﺳﺮﻣﺎﯾﻪ ﭘﺬﯾﺮان ﺗﺄﻣﯿﻦ ﻣﺎﻟﯽ ﺟﻤﻌﯽ ﻣﺒﺘﻨﯽ ﺑﺮ ﺑﺪﻫﯽ اﺳﺖ. ﺑﺎ ﺗﻮﺟﻪ ﺑﻪ ﭘﯿﭽﯿﺪﮔﯽ ارزﯾﺎﺑﯽ رﯾﺴﮏ، ﺑﻬﺘﺮﯾﻦ ﻣﻌﻤﺎري ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﺑﺎ ﺧﺎﺻﯿﺖ ﺗﻨﻈﯿﻢ ﻧﺮون ﻻﯾﻪ ﭘﻨﻬﺎن، اﻟﮕﻮرﯾﺘﻢ ﭘﺮﺳﭙﺘﺮون ﭼﻨﺪ ﻻﯾﻪ ﺑﺮاي ﺷﺒﯿﻪ ﺳﺎزي اﻧﺘﺨﺎب ﺷﺪ. ﺟﺎﻣﻌﻪ آﻣﺎري اﯾﻦ ﭘﮋوﻫﺶ، اﻃﻼﻋﺎت ﻣﺎﻟﯽ ﭘﺮوﻧﺪه اﻋﺘﺒﺎري/ﺗﺴﻬﯿﻼﺗﯽ ﮐﻠﯿﻪ ﻣﺸﺘﺮﯾﺎن )506 ﻣﻮرد( ﯾﮑﯽ از ﺑﺎﻧﮏ ﻫﺎي ﮐﺸﻮر ﻣﺮﺑﻮط ﺑﻪ ﺳﺎل 97-98 اﺳﺖ. ﺑﻪ ﻣﻨﻈﻮر ﻣﻌﻨﺎداري راﺑﻄﻪ ﺷﺎﺧﺺ ﻫﺎي اﺳﺘﺨﺮاج ﺷﺪه از ﻧﻤﻮﻧﻪ ﺑﺎ ﻣﺘﻐﯿﺮ ﺧﺮوﺟﯽ ﻣﺪل )ﻧﮑﻮل و ﻋﺪم ﻧﮑﻮل( اﻋﻀﺎي ﻧﻤﻮﻧﻪ ﺗﻮﺳﻂ آزﻣﻮن رﮔﺮﺳﯿﻮن ﺳﻨﺠﯿﺪه ﺷﺪ. ﺑﺪﯾﻦ ﺗﺮﺗﯿﺐ ﺗﻌﺪاد13ﺷﺎﺧﺺ ﺑﻪ ﻋﻨﻮان ﺑﺮدار ورودي ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﺑﺎ ﺳﻪ ﻻﯾﻪ ﭘﻨﻬﺎن در دو ﮔﺮوه ﻧﮑﻮل و ﻋﺪم ﻧﮑﻮل وارد ﻣﺪل ﮔﺮدﯾﺪ. ﺑﺮ اﺳﺎس ﻧﺘﺎﯾﺞ ﺷﺒﯿﻪ ﺳﺎزي، ﻣﺪل ﭘﯿﺸﻨﻬﺎدي ﺗﻮاﻧﺴﺖ ﺑﺎ ﺧﻄﺎي ﮐﻤﺘﺮ و دﻗﺖ ﭘﯿﺶ ﺑﯿﻨﯽ ﺑﺎﻻﺗﺮ )94,1( وزن ﻫﺮﯾﮏ از ﺷﺎﺧﺺ ﻫﺎي ورودي ﺑﻪ ﺷﺒﮑﻪ را ﻣﺤﺎﺳﺒﻪ ﮐﻨﺪ. ﻫﻤﭽﻨﯿﻦ ﺿﺮﯾﺐ ﺗﻌﯿﯿﻦ ﺑﺮاي داده ﻫﺎي آﻣﻮزﺷﯽ ﺑﺮاﺑﺮ )0,88(، آزﻣﺎﯾﺶ ﺑﺮاﺑﺮ )0,94( و ارزﯾﺎﺑﯽ ﺑﺮاﺑﺮ )0,84( ﺑﺪﺳﺖ آﻣﺪ ﮐﻪ اﯾﻦ اﻣﺮ ﻧﺸﺎن دﻫﻨﺪه ﺗﻮاﻧﺎﯾﯽ ﺑﺮازش ﺑﺎﻻي ﻣﺪل ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﭘﯿﺸﻨﻬﺎدي اﺳﺖ. ﯾﺎﻓﺘﻪ ﻫﺎي ﭘﮋوﻫﺶ ﻧﺸﺎن داد، از ﻣﯿﺎن ﺷﺎﺧﺺ ﻫﺎي ورودي، درآﻣﺪﺧﺎﻟﺺ، ﺑﺎ وزﻧﯽ ﻣﻌﺎدل 0,163، ﻣﯿﺎﻧﮕﯿﻦ ﺣﺴﺎب ﺟﺎري ﺑﺎ وزﻧﯽ ﻣﻌﺎدل 0,123 ﺑﻪ ﻣﺮاﺗﺐ از اﻫﻤﯿﺖ ﺑﯿﺸﺘﺮ و ﺷﺎﺧﺺ ﺳﺎﺑﻘﻪ ﺗﺤﺼﯿﻼت ﺑﺎ وزﻧﯽ ﻣﻌﺎدل 0,053 از اﻫﻤﯿﺖ ﮐﻤﺘﺮي در ﮔﺮوه ﻋﺪم ﻧﮑﻮل ﺷﺪه ﺑﺮﺧﻮردار اﺳﺖ

The purpose of this study is predict and design Credit risk model for debut crowdfunding .According, the complexity of the risk assessment the best neural network architecture with Customize hidden layer neurons selected Multilayer perceptron algorithm for simulation. The statistical population of this study is the financial information of credit / loan file of all customer (506 cases) one of the banks of the country for the year 1997-98. In order to show the significant relationship the extracted indices of the sample and the model output variables (non-default and default), the sample member tested by regression.Thus, thirteen indices entered to the model neural network input vector with three hidden layers in non-default and default groups. In the simulation results, the proposed model was able to optimize the weights of each of the inputs to the network with lower prediction error and 94.1% efficiency .also the average error absolute value obtained for training data (0.88), test data (0.94) and evaluation data (0.84) indicating high capability of the proposed model. According to the research Results, among the indices, income, 0.163 weight, Current Account weight 0.123 are more important, but “degree of education of education” 0.053 are less important in the non-defaulted group.