WorksMathematics
Markov potential priors for prediction in Gamma models
Abstract
We consider Bayesian prediction in independent Gamma scale models with arbitrary known positive shape parameters. When the dimension is at least three, we construct a shrinkage prior toward any prespecified positive scale vector. The resulting predictive density strictly dominates the Jeffreys predictive density under Kullback–Leibler loss and is minimax. The prior is obtained by weighting the Jeffreys prior with a Markov potential. The transition is determined by the beta–gamma decomposition. The domination result is nonasymptotic.