Maximum likelihood prediction (Q1070691)

Let X and Y denote (generally dependent) random vectors with joint probability density f(x,y;\(\beta)\), where \(\beta\) is a parameter. The problem is to predict the unobservable (either future or past or missing) value of Y, having observed X. The predictive likelihood function is then \(L(y,\beta;x)=f(x,y;\beta).\) Suppose \(Y^*\) and \(\beta^*\) are statistics (functions of X) for which \(L(y^*,\beta^*,x)=\sup_{(y,\beta)}L(y,\beta;x).\) Then \(Y^*\) is the maximum likelihood predictor of Y and \(\beta^*\) is the predictive maximum likelihood estimator of \(\beta\). The approach is thus non- Bayesian. The main application is to the prediction of higher order statistics from lower ones in type II censored random samples.