Publication: PROPUESTAS PARA LA PREDICCIÓN DEL SOBREENDEUDAMIENTO EN HOGARES DE CHILE MEDIANTE EL USO DE UN MODELO HÍBRIDO QUE MEZCLA "ARTIFICIAL NEURO FUZZY INFERENCE SYSTEM" Y MODELO PROBIT
Loading...
Date
2018
Authors
ASTUDILLO FINGERHUT, NICOLE KARINA
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Este trabajo propone un mejoramiento en la predicción del sobreendeudamiento delos hogares en Chile, utilizando para ello un innovador modelo hibrido basado en lasmetodologías Adaptative Neuro Fuzzy Inferences System (ANFIS) y Probit.Para comparar los resultados obtenidos por el modelo propuesto, este trabajo incluye laaplicación de dos técnicas de aprendizaje inteligente ampliamente utilizadas en la literatura;las Redes Neuronales (ANN), en particular el Perceptrón Multicapa (MLP) y Suppor VectorMachine (SVM).Para cada modelo utilizado se realiza una optimización de sus parametros más relevantespara la obtención de resultados. Para el caso de la red MLP se optimiza el número deneuronas de la capa escondida, la tasa de aprendizaje y el momentum. Para el modelo SVM,se obtienen los parametros optimos de costos C y de kernel . Por último, para el caso delmodelo ANFIS y ANFIS-Probit se determinan los parametros de funciones de membresía.Los datos utilizados corresponden a la última versión de la Encuesta Financiera deHogares 2014 elaborada por el Banco Central y cuyas observaciones utilizadas en estainvestigación alcanzan los 3425 datos. El procesamiento de datos abarca el escalamiento yla aplicación de la técnica Smote para abordar el problema de desbalance de clases.Los resultados obtenidos permiten indicar que el modelo propuesto es superior a laslos modelos benchmark en términos de Tasa de precisión y Tasa de precisión promedioademás del Error tipo I.
The increase in debt levels of families in dierent parts of the world has attracted theattention of local and global organizations dedicated to the prevention of financial risks,and has intensified the interest in developing early detection methods for over-indebtednessin the population. The present work proposes a hybrid model of Adaptative Neuro FuzzyInferences System (ANFIS) for the prediction of household over indebtedness, based ona statistical technique and Neuro Fuzzy. The proposed model was compared with theProbit, Multilayer Perceptron (MLP) and Support Vector Machine (SVM) models. Themost relevant parameters for the performance of each technique are optimized, and wemanage the data balance problems through the Smote oversampling technique. We usedata obtained from the Financial Household Survey of the Central Bank (EFH) 2014 ofChile. The results show that the proposed model has a significantly better performance than the reference models in terms of the correct classification rate, the average correctclassification rate and the type I error. Consequently, this work provides an innovativeunderstanding of the problem of over-indebtedness of households that can be very usefulfor dierent governmental entities focused on preventing excessive indebtedness andmaintaining financial stability.
The increase in debt levels of families in dierent parts of the world has attracted theattention of local and global organizations dedicated to the prevention of financial risks,and has intensified the interest in developing early detection methods for over-indebtednessin the population. The present work proposes a hybrid model of Adaptative Neuro FuzzyInferences System (ANFIS) for the prediction of household over indebtedness, based ona statistical technique and Neuro Fuzzy. The proposed model was compared with theProbit, Multilayer Perceptron (MLP) and Support Vector Machine (SVM) models. Themost relevant parameters for the performance of each technique are optimized, and wemanage the data balance problems through the Smote oversampling technique. We usedata obtained from the Financial Household Survey of the Central Bank (EFH) 2014 ofChile. The results show that the proposed model has a significantly better performance than the reference models in terms of the correct classification rate, the average correctclassification rate and the type I error. Consequently, this work provides an innovativeunderstanding of the problem of over-indebtedness of households that can be very usefulfor dierent governmental entities focused on preventing excessive indebtedness andmaintaining financial stability.
Description
Catalogado desde la version PDF de la tesis.
Keywords
ANFIS , MODELO HIBRIDO , MODELO PREDICTIVO , SOBREENDEUDAMIENTO EN HOGARES