Numerical method for estimating multivariate conditional distributions (Q2488388): Difference between revisions

It is assumed that the response \(y\) depends on regressors \(x\) through a model \(y=\mu_\omega(x)+A^{-1}_\omega(x)z\), where \(z\) is a random vector with a fixed ``standard'' distribution \(G_\omega\), \(\omega\) is an unknown parameter, and \(\mu\) and \(A\) are known up to \(\omega\). The training sample contains \((y_i,x_i)\) which relates to i.i.d. copies of \(z_i\). A maximum likelihood estimation procedure for \(\omega\) is discussed. A stochastic global optimisation algorithm is described for searching the likelihood maximum. The obtained estimates are used for prediction of \(y\). Efficiency of forecasts obtained by this technique is compared with linear regression, neural network and kernel-based distribution estimation predictions. Simulation results and application to economic data are presented.