Paper shows that, under assumption that the single forecasts which enter the combination are unbiased, imposing some constraints on coordinates of <em>M </em>-estimator (of corresponding regression coefficients) leads to a gain in the asymptotic variance of one-step forward prediction evaluated by means...

Consistency, asymptotic representation and asymptotic normality of the least trimmed squares estimator of regression coefficients is derived in the framework with random carriers. Also a result describing sensitivity to the deletion of an observation is given.