A team of researchers at the University of Alcalá (UAH) has developed a system based on neural networks to accurately detect and predict a financial crisis, as published in the magazine Omega. The system of macroeconomic data and information provided by banks.
A multidisciplinary research group involving Sancho Salcedo and Jose Antonio Portilla, researchers and professors in the Department of Signal Theory at the University of Alcala, has developed a system based on techniques of nonlinear neural networks that can predict from information from a database of macroeconomic and banking, financial crisis.
“We had a fairly large database, with different parameters related to macroeconomic parameters over a hundred countries, compiled from 1981 to 1999. On the other hand, we had the data on whether the country had entered the systemic financial crisis in those years or not selected. Using neural network techniques present data for the network to learn and we have a system capable of predicting a financial crisis with a reliability of 91%, “says Salcedo.
The techniques presented in this work were compared with alternative methodologies, such as rought sets, which achieve an accuracy rate of 86% and were the best results so far. This work has been published in the prestigious international journal Omega, an improvement of existing systems so far because it extracts more information from the database.
Fundamentally, the proposed system uses a type of neural networks called product units and radial basis functions, hybrid logistic regression model. The original system for regression, with another application, was developed by the group of Professor César Hervás, University of Cordoba. The system produces a series of nonlinear models involving macroeconomic variables. The interpretation of the results was conducted by Professor Maria Jesus Segovia, part of the research group Mathematical Methods applied to Actuarial Science, as well as Alicia Sanchis, researcher linked to the Bank of Spain.
“Overall, not very interpretable models are obtained by high non-linearity, ie, they are more like a black box model, but results in probability of error on the test sample of say we are right in a high percentage of occasions, “said Sancho Salcedo.
And the current crisis, “could have predicted? Definitely not. As explained by Professor Salcedo, the current world crisis is completely different from others that have happened in the past. “The information contained in the database with which we work is not enough to predict a crisis like the present. Within a time, once to take and analyze the facts of this situation, we may have new information relevant to improving the system. “ And is that one of its features is that it is able to ‘re-train’ and learn from the new data are added
PA Gutierrez, MJ Segovia-Vargas, S. Salcedo-Sanz, C. Hervas-Martinez, A. Sanchis, JA Portilla-Figueras, F. Fernandez-Navarro, “Hybridizing logistic regression with RBF networks and product unit for Accurate detection and prediction of Banking Crises”
Omega 38 (5): 333-344, 2010.
Source: Universidad de Alcala
|Category: Information Technology||Tags: macroeconomic, nonlinear neural networks|