Shrinkage methods have traditionally been applied in prediction problems. In this article we develop a shrinkage method (mean subset) that forms an average of regression coefficients from individual subsets of the explanatory variables. A Bayesian approach is taken to derive an expression of how the coefficient vectors from each subset should be weighted. It is not computationally feasible to calculate the mean subset coefficient vector for larger problems, and thus we suggest an algorithm to find an approximation to the mean subset coefficient vector. In a comprehensive Monte Carlo simulation study, it is found that the proposed mean subset method has superior prediction performance than prediction based on the best subset method, and in some settings also better than the ridge regression and lasso methods. The conclusions drawn from the Monte Carlo study is corroborated in an example in which prediction is made using spectroscopic data.