On the selection of subset bilinear time series models: a genetic algorithm approach (Q1861633): Difference between revisions

The authors consider a bilinear time series model of the form \[ x_t=\phi_0+\sum_{i=1}^p\phi_i x_{t-i}+e_t-\sum_{j=1}^q\vartheta_je_{t-j} +\sum_{i=1}^{b_p}\sum_{j=1}^{b_q}\beta_{ij}x_{t-i}e_{t-j}, \] where \(e_t\) are i.i.d. Gaussian (\(0,\sigma^2\)) errors. The problem is to determine the specification of the model, i.e., the coefficients \(\phi\), \(\psi\), \(\beta\) which should be nonzero. A genetic algorithm is described in which specifications are considered as ``chromosomes'' with abilities to reproduction, mutation and crossover. A population of such chromosomes proliferate and the best fit ones survive. The authors use Akaike's AIC and BIC, and Schwartz's SBC as fitness criteria. Illustrative simulation examples and real sunspot numbers data from 1700-1983 are considered.