This paper evaluates linear and non-linear forecast-combination methods. Among the non-linear methods, we propose a nonparametric kernel-regression weighting approach that allows maximum flexibility of the weighting parameters. A Monte Carlo simulation study is performed to compare the performance of the different weighting schemes. The simulation results show that the non-linear combination methods are superior in all scenarios considered. When forecast errors are correlated across models, the nonparametric weighting scheme yields the lowest mean-squared errors. When no such correlation exists, forecasts combined using artificial neural networks are superior.