In examining MJO skill in the SubX reforecasts, we compare the real-time multivariate MJO (RMM) index to two new metrics: (1) a time-extended empirical orthogonal function (EOF) using velocity potential and (2) wavenumber-frequency filtered fields. The time-extended EOF is three-dimensional (time-lag, latitude, longitude) and seasonally varying, which allows it to better capture the differences in MJO structure between summer and winter. In addition, because the time-extended EOF excludes the influence of high-frequency signals such as Kelvin waves it is more predictable than the RMM index. Compared to the global MJO indices, the wavenumber-frequency filtering approach is better able to depict regional variations in the skill of the models. We examine how MJO forecast skill relates to systematic biases and errors in MJO phase speed and amplitude in the different models and MJO indices. We also examine the temporal variability of skill and compare high and low skill periods in each of the models. Lastly, differences in MJO skill between the S2S models are related to differences in the ability of the models to predict mid-latitude teleconnections and tropical cyclone activity at S2S time scales.