A suite of methods that have been proposed for statistical post-processing of ensemble forecasts based on historical verification data (i.e. ensemble-MOS methods) are compared with each other, and with direct probability estimates using ensemble relative frequencies, in the idealised Lorenz '96 setting. The three most promising methods are logistic regressions predicting probabilities associated with selected quantiles, ensemble dressing (a kernel density estimation approach), and linear regressions with non-constant prediction errors that depend on the ensemble variance.