Adaptive Neuro-Fuzzy Inference Systems with Heteroscedastic Errors for

Abstract

This paper proposes a new kind of non-linear hybrid model. In the proposed model, mean non-linearity is represented by using an adaptive neuro-fuzzy inference system (ANFIS) whereas variance is represented using a conditional self-regressive heteroscedastic component. The mathematical formula for this type of model is shown and a method to estimate it is proposed. In addition, a specification strategy is developed for the proposed model, based on a battery of statistical soft transaction regression (STR) tests and on verosimility radius testing. As a case study, the IBM stock closing price series dynamics were modeled, which is commonly used as a benchmark in the literature on time series. Results indicate that the model developed represents the dynamics of the studied series better than other models with similar characteristics.