For all three observed Indian rainfall datasets, the four wettest months extend from June to September with maximum rainfall in July (Figure 2a). The CMIP models show a range of behaviors. Eleven out of 24 CMIP3 models (Figure 2a) and 11 out of 23 CMIP5 models (Figure 2c) correctly simulate peak rainfall in July, with several models with peak rainfall delayed by one month. Three CMIP3 models have large timing biases: the mpi_echam5 and csiro_mk3.5/ipsl-cm4 peak two months early and late respectively. There is also considerable spread in the amplitude of the seasonal cycle. While rainfall in the peak monsoon month ranges from 7.2 to 9.1 mm/d across observations, both CMIP3 and CMIP5 models range from ∼3 to 10 mm/d. The multi-model means (±1 standard deviation) for CMIP3 and CMIP5 models are similar with a maximum of ∼6 ± 2 mm/d in August, indicating that in general the monsoon rainfall in the models is too weak. For a few models (particularly for CMIP3), there is too much rainfall outside of the monsoon season (e.g., cnrm_cm3 and inmcm3_0).
Figure 2. Seasonal cycle of the (a, c) Indian and (b, d) Australian land-restricted rainfall for observed and CMIP3 (Figures 2a and 2b) and CMIP5 (Figures 2c and 2d) models over. Models (names in black) and observations (names in red) are sorted according to the average monsoon rainfall amount (JJAS rainfall for Indian monsoon and DJFM rainfall for Australian monsoon). The top row shows the multi-model mean (names in blue) of CMIP3 or CMIP5 models for Indian and Australian rainfall. Internal numbers show the maximum rainfall (mm/d) in the month of greatest rainfall.
Download figure to PowerPoint
 Over Australia maximum observed rainfall occurs in February, ranging between 5.9 to 8.6 mm/d across the three observational datasets (Figure 2b) with most rainfall from December to March. Again there is a considerable range in monsoon strength across the models (2 to 11 mm/d), although the multi-model means (7.5 ± 3/7.0 ± 3 mm/d for CMIP3/CMIP5) lie within the observational range. The giss_aom and ipsl_cm4 CMIP3 models have essentially no monsoon season. Most CMIP3 and CMIP5 models simulate maximum rainfall in the correct month with a few of models peaking one month early and others with an overly long rainy season (e.g., bccr_bcm2_0).
 In general, for both CMIP3 and CMIP5 and for both regions the range in monsoon strength across the models is about the same. However, there is a clear overall improvement in the seasonality of both monsoons from CMIP3 to CMIP5, with most CMIP5 models better simulating the monsoon timing and very low rainfall rates outside of the monsoon season.
3.2. TBO Transitions Assessment
 A Monte Carlo technique is used to assess the significance of enhanced predictability associated with the TBO. In particular we assess the success rate of the four transitions that are thought to be important for the TBO tendency starting from a given year t: (1) Successful Indian-Indian out-of-phase transition is defined as IMRI(t) > 0 and IMRI(t + 1) < 0 or IMRI(t) < 0 and IMRI(t+ 1) > 0; (2) Successful Australian-Australian out-of-phase transition is defined as AMRI(t) > 0 and AMRI(t + 1) < 0 or AMRI(t) < 0 and AMRI(t+ 1) > 0; (3) Successful Indian-Australian in-phase transition is defined as IMRI(t) > 0 and AMRI(t) > 0 or IMRI(t) < 0 and AMRI(t) < 0; and, (4) Successful Australian-Indian out-of-phase transition is defined as AMRI(t) > 0 and IMRI(t + 1) < 0 or AMRI(t) < 0 and IMRI(t + 1) > 0.
 To determine the observed predictability associated with a given transition, we count the number of successful transitions in the timeseries (Figures 1b and 1c), relative to the total number of possible successful transitions. This is then repeated 100,000 times by randomly resampling the observed timeseries (with replacement). The predictability of the monsoon resulting from a given transition is considered enhanced if the observed or simulated percentage of successful transitions is significantly higher than the median of the randomized distribution. Moreover, the enhanced predictability is considered significant if the observed or simulated percentage of successful transitions lies in the upper decile of the randomized distributions (i.e., there is only a 10% probability of getting the enhanced predictability by chance).
 In the IMRI-AIR and AMRI-AWAP, only two of the four transitions show a significantlyenhanced predictabilityover the random distribution: the Indian-Indian out-of-phase transition and the Indian-Australian in-phase transition, where theenhanced predictability from the median is 8.8% (p ∼ 0.1) and ∼15% (p ∼ 0.001), respectively (Figure S1 in the auxiliary material). This means that if the Indian monsoon is anomalously strong (weak) in a given year there is a ∼59% probability that the subsequent Indian monsoon will be anomalously weak (strong) and a ∼65% probability that the subsequent Australian monsoon will be anomalously strong (weak).
 To test the sensitivity of these results to the datasets, we compute predictability based on different datasets and time periods: CMAP (1979–2008) and GPCC (1901–2010) over the same spatial land areas (Figure 1a). Indices from both datasets show significant enhanced predictability(p ∼ 0.1) of 27.6% and 13.8%, respectively, for the Indian to Australian transition. However, for the Indian-Indian out-of-phase transition, only GPCC shows anenhanced predictabilityof 6.4% (p ∼ 0.12). Given the relative shortness of the CMAP datasets, the lack of significant results might arise from multi-decadal variability of the TBO. To examine this, we divided the IMRI-AIR into a CMAP-like period (1979–2008) and pre-CMAP period (1871–1978). The pre-CMAP period shows significantlyenhanced predictability(10.3%, p ∼ 0.1). The predictability for over the later CMAP-like period is not significant. This suggests that the biennial tendency for the Indian-Indian out-of-phase transition has deceased in the past 30-yrs, consistent with the recent weakening of the El Niño-Indian monsoon relationship [Ummenhofer et al., 2011; Meehl and Arblaster, 2011].
 Figure 3 shows the enhanced predictability for the different observational datasets and selected CMIP3 and CMIP5 models for the four transitions. We only show models for which at least one ensemble member shows a significantly enhanced predictability in any of the transitions. Twelve CMIP3 and 15 CMIP5 models have at least one ensemble member with significantly enhanced predictabilityfor the Indian-Australian in-phase transition (Figure 3a), consistent with the observations. For the CMIP5 models the range of enhanced predictability is relatively small (5%–15%, for significant results only). As such the models generally underestimate the observed enhanced predictabilityof ∼15%. For the CMIP3 models the range is larger (5–35%). For the models with multiple ensemble members, many show contrasting results for different members, indicating a substantial multi-decadal variability in the efficacy of this transition. Nevertheless, when considering concatenated ensemble members, the Indian- Australian transition hasenhanced predictability in 11 of 24 (14 of 23) CMIP3 (CMIP5) models.
Figure 3. Percentage enhanced predictabilityfor the (a) Indian-Australian, (b) Indian-Indian, (c) Australian-Indian and (d) Australian-Australian transitions for observations and CMIP3 and CMIP5 models for which at least one ensemble member for that model and for at least one of the transitions shows a significant increase in predictability. Circles represent individual ensemble members (marked in red are significant, p < 0.1). Bars represent the multi-ensemble mean percentageenhanced predictabilityfor each model with yellow indicating significant changes. The multi-ensemble mean predictability was calculated by concatenating time series for all ensemble members prior to Monte Carlo resampling.
Download figure to PowerPoint
 As noted above, the only other transition that provides significantly enhanced predictabilityusing the long-term reference observations is the Indian-Indian out-of-phase transition, although it is mostly associated with the earlier part of the record (see discussion above). For this transition (Figure 3b), seven CMIP3 and five CMIP5 models show enhanced predictability for at least one ensemble member. In the case of models with multiple ensemble members, only a small number of members are significant (e.g., CNRM_CM5, GISS_E2_H and NorESM_M). As such, only four CMIP3 and two CMIP5 models show enhanced predictabilitywhen considering the multi-ensemble concatenation for each model.
 For the Australian-Indian out-of-phase transition (Figure 3c), one of the longer observational timeseries suggests a multi-decadal period when there wasenhanced predictability associated with this transition. Five CMIP3 and only one CMIP5 models have ensemble members with significantly enhanced predictability, despite conflicting results within each ensemble set. Surprisingly, a large number of models actually show significant negative predictability, in particular when considering concatenated ensemble members. Negative predictability suggests that a strong (weak) Australian monsoon would tend to be followed by a strong (weak) Indian monsoon. Such behavior is not found in the observations, and might be related to bias in the simulated ENSO seasonal cycle, e.g., where La Niña events tend to persists too long this would lead to a strong Australian monsoon followed by a strong Indian monsoon (N. C. Jourdain et al., The Indo-Australian monsoon and its relationship to ENSO and IOD in reanalysis data and the CMIP3/CMIP5 simulations, submitted toClimate Dynamics, 2012). Finally for the Australian-Australian out-of-phase transition (Figure 3d), for which there is no observational evidence of enhanced predictability, eight of the CMIP3 and five of the CMIP5 models have ensemble members with significantly enhanced predictability.
 While the impacts of the monsoon are primarily over land, the teleconnections associated with global modes of variability can involve larger domains. To assess the effect of the monsoon in a larger context, we also derived similar rainfall indices over extended Indian and Australian regions (including the oceanic regions, Figure 1a, see methods) used in previous analysis of the TBO [e.g., Meehl and Arblaster, 2002]. Applying the Monte Carlo analysis to these indices for both CMIP3 and CMIP5 models, we find that for the Indian-Australian transition, most models show significantenhanced predictability independent of the index definition (Figure S2). For the other transitions there is little consistency with regards to predictability across the models. For the Australian-Indian transition, fewer models show the negative predictability seen with the land-only based indices. In addition for the Australian-Australian transition there are more CMIP3 models that exhibitenhanced predictabilityusing the more inclusive index. As such, overall we do not find that the sequence of events that combine to make up the TBO is more obvious when examining a larger land-ocean area.