Since the Madden–Julian oscillation (MJO) is a major source for tropical and extratropical variability on weekly to monthly timescales, the intrinsic predictability of its global teleconnections is of great interest. As the tropical diabatic heating associated with the MJO ultimately drives these teleconnections, the variability in heating among ensemble forecasts initialized from the same episode of the MJO will limit this predictability. In order to assess this limitation, a suite of 60 d ensemble reforecasts has been carried out with the ECMWF forecast model, spanning 13 starting dates from 1 November and 1 January for different years. The initial dates were chosen so that phases 2 and 3 of the MJO (with anomalous tropical heating in the Indian Ocean sector) were present in the observed initial conditions. The 51 members of an individual ensemble use identical initial conditions for the atmosphere and ocean. Stochastic perturbations to the tendencies produced by the atmospheric physics parameterizations are applied only over the Indian Ocean region (50–120∘ E). This guarantees that the spread between reforecasts within an ensemble is due to perturbations in heat sources only in the Indian Ocean sector. The point-wise spread in the intra-ensemble (or error) variance of vertically integrated tropical heating Q is larger than the average ensemble mean signal even at early forecast times; however the planetary wave (PW) component of Q (zonal waves 1–3) is predictable for 25 to 45 d, the time taken for the error variance to reach 50 % to 70 % of saturation. These scales never reach 90 % of saturation during the forecasts. The upper-level tropical PW divergence is even more predictable than Q (40 to 50 d). In contrast, the PW component of the 200 hPa Rossby wave source, which is responsible for propagating the influence of tropical heating to the extratropics, is only predictable for 20 to 30 d. A substantial ensemble spread of 300 hPa meridional wind propagates from the tropics to the Northern Hemisphere storm-track regions by days 15–16. Following the growth of upper-tropospheric spread in planetary wave heat flux, the stratosphere provides a feedback in enhancing the error via downward propagation towards the end of the reforecasts.