The CSIRO ESM interactively couples the Conformal-Cubic Atmospheric Model (CCAM; McGregor and Dix, 2008), dynamic ocean model (Thatcher et al., 2015), sea-ice model (Gordon et al., 2002; O’Farrell, 1998) and sophisticated new generation land-surface model referred to as the CSIRO Atmosphere Biosphere Land Exchange (CABLE; Kowalczyk et al., 2013). The model employs a common reversibly staggered grid (McGregor, 2005), eliminating the coupling overhead associated with message passing among model components. This approach makes the simulations computationally economical since it negates the need for interpolation in every model time step commonly applied in other coupled models to reconcile grid type and resolution differences. In addition, the CSIRO ESM includes a prognostic aerosol scheme due to Mitchell et al. (1995), applied consistently with the emission inventories and radiative forcing specifications of the latest Coupled Model Intercomparison Project (CMIP; Taylor et al., 2012). It also integrates a dynamic river routing scheme of Arora and Harrison (2007) which feeds fresh water to the ocean and surface water. The model furthermore incorporates biogeochemical knowledge and elements of the terrestrial carbon cycle.

The study uses 15 yrs. (2000–2014) retroactive ensemble simulations built as a function of atmospheric states and ocean nudging; each integration runs for 6 months in length. This model configuration offers a better description of uncertainties that may arise from the initial and boundary forcings. In this regard, the CSIRO ESM is nudged, at its surface, to the Multi-Model Ensemble (MME) mean and variablity amplitude of the SSTs using spectral nudging (Thatcher and McGregor, 2009, 2010) to minimize the potential climate drift (large and unrealistic departures between the RCM and its driving model/s’ fields) (Omrani et al., 2012). This is a necessary step, to contain CSIRO ESM’s generated SSTs drift which will affect the RCM outputs. The MME SST forcing is obtained from North America (NMME) through the IRI data library, University of Pretoria Statistical Model (Landman et al., 2014), and SINTEX-F2v. of Japan Agency for Marine-Earth Science and Technology (JAMSTEC; Doi et al., 2016). The 18 CRCM ensemble members for hindcasts are created from the six scaled time-lagged initial conditions (ICs) in conjunction with the three SST states. The atmospheric ICs are acquired from the NCEP (National Centers for Environmental Prediction), Department of Energy (DOE) Atmospheric Model Intercomparison Project (AMIP) II Reanalysis (R2) data set (Kanamitsu et al., 2002). Each of these time-lagged perturbations are prescribed with the time varying CO₂ and Ozone boundary conditions.

In this experiment, the CSIRO ESM is run at C192 quasi-uniform (approximately about 50 km) global horizontal resolution and 25 vertical sigma layers. With a self-constraining stretched-grid paradigm, the CSIRO ESM runs, further, constrain the C192 variable spatial resolution of the same model by applying a Schmidt factor of 0.1523, centered over southern Africa (28°S, 25°E). This yields about 8 km finer horizontal resolution (CRCM) over continental Africa whilst maintaining coarse resolution elsewhere. The use of high-resolution (50 km) coupled forcing is found to be essential to maintain numerical stability and achieve balanced simulations of CRCM consistent with the driving CSIRO ESM. More description of CSIRO ESM can be found in Beraki et al. (2020) which elucidated the atmosphere–ocean-sea-ice interactions in response to different climate forcings and how these interactions affect the Southern Hemisphere’s circulation variability. Furthermore, CRCM is currently being experimentally used for seasonal prediction and its applications in renewable energy and precision Agriculture in South Africa. This study is the first effort to cover the process approach evaluation of this model set-up for seasonal timescales. Noteworthy, we are not aware of other initiatives that use similar modeling paradigm reported here for seasonal prediction over the African continent which has revealed a great potential for climate applications in an operational environment.

The evaluation is done on several months lead time hindcasts, generated from the November initialization. To minimize the loss of skill with the lead time (Brum and Schwanenberg, 2022), the seasons under consideration consist of the rolling 3 months seasons beginning with November. November was chosen for initialization because the main Southern Africa’s rainy season starts in November and ends in February, with a tendency of extending up to April (Macron et al., 2014; Pohl et al., 2009, 2018; Ullah et al., 2023; Vigaud et al., 2012). Madagascar, for example, experiences the wettest period between November and April (Randriatsara et al., 2022). Over Southern Africa, March to May (autumn) and September to November (spring) are the transitional seasons (Wang et al., 2021) that are not used in this study. November initialization is, furthermore, explored because it is aligned to Mamalakis et al. (2021) study of the ITCZ using a suite of Coupled Model Intercomparison Project, version 6 (CMIP6) global models (Eyring et al., 2016). The extreme North/South position of the ITCZ corresponds to peak austral winter/summer seasons (Zhang et al., 2021) respectively. These extreme positions provide the necessary seasonal boundary positions of the ITCZ over the African tropics and forms part of the reasons why summer (November) initialization is evaluated to identify the extreme southern position of the ITCZ. Figure 1 provides the schematic summary of the CSIRO ESM and CRCM model set-up configuration and interactions as applied in this research.

The standard evaluation procedure is applied; the gridded observational/reanalysis datasets and model (CSIRO ESM and CRCM) hindcasts, for the period 2000–2014, were used to identify the seasonal ITCZ position and statistically compared the resulting positions. According to MacLachlan et al. (2015), a 14 years hindcasts is long enough to develop model climatology. A precipitation-based centroid method by Adam et al. (2016a) was used to identify the ITCZ position during El-Niño and La-Niña events. This method has recently been applied by Mischell and Lee (2022) to study the relationship between zonal ITCZ position and meridional temperature contrast. The maximum precipitation method was used to identify the ITCZ’s climatological position by identifying the latitude of maximum precipitation along each longitude as applied by Liu et al. (2020).

The observational datasets used include the gridded (0.5° × 0.5° grid) Climate Research Unit gridded Time Series version 4 (CRU TS v4; Harris et al., 2020) rainfall data and the Climate Hazards Group Infrared Precipitation with Stations version 2 (CHIRPS2.0) data at (0.05° × 0.05°) grid resolution (Funk et al., 2015). The (0.25° × 0.25°) European Centre for Medium-Range Weather Forecasts (ECMWF) fifth generation reanalysis (ERA5; Hersbach et al., 2020) dataset was also used.

CHIRPS, CRU and ERA5 datasets were used due to their value in model evaluation and interoperability as reflected by the past studies (Endris et al., 2021; Hassler and Lauer, 2021; Yimer et al., 2022). Li et al. (2021), for instance, applied them in their study on the ability of the CMIP6 models to simulate temperature and precipitation extremes. Furthermore, ERA5 dataset strongly correlates with CHIRPS & CRU (Gleixner et al., 2020) and it is also the best performing re-analysis data for statistical downscaling using analogue methods for meteorological variables (Horton, 2022). CHIRPS dataset has been proven by Burton et al. (2018) to be the best performing dataset for tropical Africa related studies. For fair comparability, we re-sampled higher resolution CHIRPS, ERA5, and Model datasets to similar (and coarser) resolution as CRU (0.5° × 0.5°) by applying the “samplegrid” operator from the Climate Data Operators (CDOs) (Schulzweida, 2023). The application of this operator does not require any interpolation schemes.

The study evaluates the model’s skill and reliability on its seasonal ITCZ simulation, including how the model performs in representing characteristics of the ITCZ such as the edge of the HC, the STJ and the stream function. To this effect, probabilistic and deterministic skill assessment metrics (Table 1) were used to assess the difference between forecasts and the reference precipitation data over the period (2000–2014). We assume that a forecast by each of the 18 ensemble members of CRCM is equally probable. A conversion of the ensemble forecast to probabilistic forecast allows use of the probabilistic validation measures such as the relative operating characteristics (ROC) curve and the Spread-error Ratio (SPR). ROC is applied to assess CRCM’s ability to discriminate amongst the above normal (AN), near normal (NN), and below normal (BN) seasonal precipitation forecasting categories. In this case AN and BN represent the upper and lower-tercile forecasts, respectively. It portrays the degree of correct probabilistic forecast discrimination (Jolliffe and Stephenson, 2011; Mason and Graham, 2002). Correct discrimination is quantified by the metric even if the forecast has a bias or calibration problems. The reliability diagram, on the other hand, evaluates the CRCM’s skill. It is used to reflect how well the forecast probabilities correspond to observation’s relative frequencies (Hagedorn et al., 2005) which is indicative of whether CRCM has over-confidence, under-confidence, conditional bias, or perfect reliability. Both the reliability diagram and ROC curves were calculated based on the CRCM’s grid-point predictions. The SPR metric is specifically chosen because it allows an assessment of variability among ensemble members and the forecast error (Wheatcroft, 2019). The benefits of SPR include its sensitivity to the reliability of a forecast without applying the event categories (Christensen et al., 2015) applied in ROC metric.

Deterministic performance metrics are also applied to quantify the correspondence between forecasts and reference data anomalies. The deterministic metrics used for evaluation include: the anomaly correlation coefficient (ACC), also known as a measure of association (Calì Quaglia et al., 2022; Liemohn et al., 2021; Ostermöller et al., 2021); the mean bias (MB, Chang et al., 2018; Liemohn et al., 2021); the root mean square error (RMSE) (Ayugi et al., 2021; Yimer et al., 2022). The Kling Gupta efficiency (KGE) introduced by Gupta et al. (2009) and later modified by Kling et al. (2012) allows an assessment of correlations between simulated and observed values, the prediction bias degree, and the variability in the modeled outputs. For example, it has been used to assess model performance in representing temporal dynamics of dry and wet bias, and dispersion (Abebe et al., 2020). The KGE values range from −0.4 to a maximum of unity (perfect skill) which reflects the model’s predictive skill in comparison to climatology (Abebe et al., 2020; Kling et al., 2012; Knoben et al., 2019). These KGE characteristics are a fit for purpose measure on whether the model can be used to explore the teleconnections between the seasonal ITCZ position and other rain bearing systems over southern Africa.

Further to the evaluation metrics explored, we have also tested CRCM’s performance on some of the main characteristics of the ITCZ. The evaluated characteristics include the edge of the HC, the multi-level (1,000 hPa to 100 hPa) stream function (Ψ) and the STJ. The identification of the HC, Ψ, and STJ was done following the python-based tropical width diagnostics software (PyTropD) as described in Adam et al. (2018), which can be consulted for further details on the relevant equations used and estimations made in identifying the edge of the HC, the STJ and mass-stream function. PyTropD has recently and successfully been applied by Menzel et al. (2024) in their study of the disconnect between the HC and STJ, and Gibson et al. (2024) for their study involving the winter precipitation change and how it relates to the shifts of the HC and the eddy driven jet. The mean absolute percentage error (MAPE) and symmetric mean absolute percentage error (SMAPE) between ERA5 reanalysis dataset and the CRCM’s outputs were calculated following Zhao et al. (2023) and Chen et al. (2017) respectively. MAPE is used here because it is part of the minimum standard metrics of the benchmarking framework as detailed by Isphording et al. (2024).

The seasonal Oceanic Niño Index (ONI) was used to identify the El-Niño and La-Niña phases. The ONI is based on 3 months running mean of SST anomalies in the Niño 3.4 Ocean region as obtained from the Extended Reconstructed SST version5 (ERSSTv5) (Huang et al., 2017). The El-Niño/La-Niña year as applied in this research, covers from November of the current calendar year (y₀) to April of the following calendar year (y₀ + 1). Therefore, NDJ (November–January) season consists of November and December from y₀ and January from y₀ + 1. DJF (December–February) consists of December from y₀ with January and February from y₀ + 1. JFM (January–March) and FMA (February–April) are from year y₀ + 1. Furthermore, a year is confirmed to be an El-Niño/La-Niña year provided the phase remains active in three of the four (NDJ, DJF, JFM, and FMA) rolling seasons. CSIRO ESM & CRCM are initialized in November and run for 6 months, which means November runs from y₀ provide predictions until April in y₀ + 1. Therefore, the active El-Niño years between 2000 and 2014 are 2002/03, 2004/05, 2009/10, and 2014/15 while the active La-Niña years are 2000/01, 2005/06, 2007/08, 2008/09, 2010/11, and 2011/12.

The figures and discussions in this section address CSIRO ESM and CRCM climatological performance on the zonal and spatial position of the ITCZ over 10–40°E longitudes using the maximum precipitation method and the question about CRCM’s ability to discriminate between above-normal (AN); near-normal (NN); and below-normal (BN) seasonal forecasting categories. The ability of CRCM to discriminate amongst the three categories is hereafter used as the basic proof of its capability to identify the ITCZ position through precipitation-based methods. The level of reliability of CRCM outputs is also discussed. All these are done for CRCM’s full climatology period (2000–2014), during the El-Niño, and La-Niña events between 2000 and 2014.

From Figures 2a,b, the CSIRO ESM and CRCM track the seasonally migrating climatological position of the ITCZ in line with both the observations (Figures 4a,b for spatial migration) and reanalysis data albeit CRCM reflecting a better performance than the CSIRO ESM. The spatial spearman correlations (at 99% significance level) between CRCM and the observation/reanalysis dataset varied from 0.66 (between CRCM and CHIRPS during JFM) to 0.91 (between CRCM and ERA5 during JFM), clearly reflected in Figure 2e as well. Figure 2a shows the zonal performance of the CSIRO ESM and CRCM against CHIRPS, CRU and ERA5 datasets for the period from November to April. CSIRO ESM and CRCM simulations can track the zonal climatological seasonal pattern of the ITCZ relative to the three test datasets (CHIRPS, CRU and ERA5). This performance is aligned to the expectation of consistency in grid-points interactions (Fox-Rabinovitz et al., 2006) between the global and regional resolutions wherein more regional details can be extracted without performance deterioration. CRCM has a northerly bias against CHIRPS and CRU while CSIRO ESM has a northerly bias against all the compared datasets (CHIRPS, CRU, ERA5 and CRCM). CRCM specifically tracks better, both zonally and spatially, with the ERA5 data than with CRU and CHIRPS possibly because ERA5 data inherently applies the same primitive equations as the weather and climate models beyond the actual observations acquired through data assimilation. For the zonal seasonal flow pattern, the southern-most latitudinal position of the ITCZ occurs during the DJF season. This seasonal shift is similarly suggested by Lauer et al. (2018), albeit for a different region.

The CRCM skill, assessed using the ROC curve against the CHIRPS dataset, found to be reasonable due to its comparable resolution to the CRCM, is shown in Figures 2, 4c. ROC is part of the metrics for evaluating climate model performance and it contributes toward the probability distribution metrics on benchmarking framework (Baker and Taylor, 2016). Figures 2, 4c shows the area under the curve (AUC) of between a minimum of 0.86 and a maximum 0.96 corresponding to JFM BN during La-Niña and DJF AN during El-Niño, respectively. These values indicate that CRCM can skillfully discriminate the AN and BN categories better than the NN (for which AUC ranges from 0.44 to 0.52) category. Furthermore, the seasonal forecast issued by CRCM is exceptionally reliable for the AN and BN categories while not reliable for the NN prediction (Figures 2, 4d). Considering that ITCZ is a high (intense) precipitation area and CRCM is skillfully reliable in high precipitation forecasting, it follows that CRCM is suitable for precipitation-based methods of identifying the ITCZ position. This capability to accurately represent high intensity precipitation events is an added value (Corney et al., 2013; Lindstedt et al., 2015) of running an ESM at a high resolution, an approach that is slightly modified by applying the stretched grid in this research. Furthermore, the reliability and ROC results imply that we can rely on the forecast for AN (even during El-Niño) or BN (during La-Niña) seasonal precipitation forecast for southern Africa which is known to experience mostly BN precipitation during El-Niño and AN precipitation during La-Niña (Guimarães Nobre et al., 2019; Lenssen et al., 2020).

A Taylor diagram compares several pairs of patterns: observations and model comparisons and tracking changes (in lead times) in model performance (Taylor, 2001). ERA5 was used as the reference dataset because of its resolution being between that of CHIRPS and CRU datasets and it has also been used to normalize CHIRPS, CRU, and CRCM standard deviations. Based on the Taylor diagram presented in Figure 2e, we note how strongly correlated ERA5, and CRCM predictions are (correlation coefficients that range from 0.89 to 0.91 at 99% confidence level); showing even similar standard deviations and root-mean square error values. On the other hand, slightly weaker correlations, higher standard deviations and higher root-mean square error values are observed for CHIRPS and CRU against CRCM predictions. In general, the Taylor diagram shows the weakening correlations between the observations/reanalysis with lead time and a high (between 1.0 and 1.6) but comparable normalized standard deviations amongst the datasets (CHIRPS, CRU, ERA5, and CRCM). The climatological error between CRCM and ERA5 is generally lower than 0.5° while it ranges from 0.75 to 1.2 between CRCM and CHIRPS/CRU. This indicates that CRCM is more aligned to ERA5 in both pattern recognition and the amplitude of variation.

In terms of the observations/reanalysis datasets correlations, CHIRPS and CRU have a much higher correlation (>0.96) than CHIRPS correlated to ERA5, and CRU correlated to ERA5 (0.47–0.81 and 0.48–0.79 correlations, respectively), detailed correlations that cover both the El-Niño and La-Niña based correlations between CRCM (CSIRO ESM) and observations/reanalysis datasets are found in Table 2. Furthermore, this is reflected in the Taylor diagram (Figure 2e) where CHIRPS and CRU are packed together and separated from ERA5 datasets. These correlations are comparable to the CHIRPS correlated to CRU results found by Steinkopf and Engelbrecht (2022) for DJF when correlating CHIRPS, CRU, and ERA5 over the tropical (between 20°S and 20°N) Africa. The climatological normalized standard deviations amongst the observations/reanalysis and CRCM are all comparable throughout the seasons (maximum difference of <0.6) while the RMSE varies between 0.4 and 1.25°.

The analysis of the CSIRO ESM and CRCM performance over different rolling seasons (NDJ, DJF, JFM, and FMA) during the positive and negative phases of ENSO is shown in Figure 3. The centroid method was used for the identification of the spatial and zonal positions of the ITCZ.

CSIRO ESM and CRCM consistently track the seasonally migrating spatial position of the ITCZ in line with both the observations and reanalysis data during the El-Niño/La-Niña ENSO phases (Figures 3a–d). This tracking and CRCM’s performance on precipitation prediction (Figures 2, 4c,d) agrees with Nicholson (2009) on the consistency in model performance on precipitation and ITCZ position’s prediction. However, CRCM reflects a tendency of dynamic error growth relative to CSIRO ESM as the lead time increases (Figures 3c,d). This dynamic error growth could possibly be enhanced by the resolution at which CRCM is downscaled because it coincides with the grey zone (Lindstedt et al., 2015) for resolving deep convection like commonly happens along the ITCZ. The spatial spearman correlations between CRCM (CSIRO ESM) and individual observation/reanalysis dataset range from 0.39 (0.37) between CRCM (CSIRO ESM) and CHIRPS/CRU (CHIRPS) during FMA to 0.89 (0.92) between CRCM (CSIRO ESM) and ERA5 during JFM (JFM and FMA). These results (detailed in Table 2) show a marginally better spatial correlation between CSIRO ESM and CRU/ERA5 while comparable correlations exist between CRCM/CSIRO ESM and CHIRPS observations. This implies that CSIRO ESM performs better than CRCM against the low-resolution verifying datasets while both CSIRO ESM and CRCM perform equally well against the high-resolution observational verifying datasets.

CRCM (CSIRO ESM) skill (as measured by KGE and presented by Figure 3e) in predicting the seasonally changing position of the ITCZ during El-Niño varies between 0.05 for CHIRPS during FMA for both CSIRO ESM and CRCM and 0.72 (0.78) for ERA5 during DJF. On the other hand, the CRCM’s (CSIRO ESM) skill (KGE) in predicting the seasonally changing position of the ITCZ during La-Niña varies between a minimum of −0.07 (−0.07) for CHIRPS during FMA and a maximum of 0.86 (0.89) for ERA5 during NDJ. These values indicate that using CRCM (CSIRO ESM) to predict the spatial ITCZ position during El-Niño/La-Niña is, on average at 0.38 (0.39), which is greater than −0.4 and therefore implies that it is a far better performance than relying on the climatological response of the ITCZ position to the ENSO phases. This consistency of skill between CSIRO ESM and CRCM is aligned to the expectation in terms of the general performance between the RCM and its driving ESM as stated by Fox-Rabinovitz et al. (2006). Once again CSIRO ESM is skillfully better than CRCM against the low-resolution verifying datasets while they are equally skillful against high-resolution verifying datasets.

CRCM (CSIRO ESM) skill is with the level of reliability, measured by SPR and presented by Figure 3f, that varies between 0.89 (1.36) for ERA5 during FMA to 2.23 (2.50) for CRU (CHIRPS) during NDJ under the positive phase of ENSO while it varies between 0.80 (1.32) for ERA5 (ERA5/CRU) during FMA to 2.57 (3.01) for ERA5 during NDJ under the negative phase of ENSO. This indicates that, on average, CRCM (CSIRO ESM) gives a slightly better performance during La-Niña than it does during El-Niño even with an element of overconfidence (under dispersion) by CRCM. CRCM overconfidence also exists on the climatological seasonal position of the ITCZ (Figure 2d), especially for forecast probabilities of 40–90% for BN precipitation during the DJF and JFM.

Multiple studies, from as far back as the early 2000s (Landman et al., 2009) to as recent as the past 5 years (Scaife et al., 2019; Tapiador et al., 2020), have shown an interest in identifying the value added by the RCMs from the global circulation models (GCMs) and earth system models on the seasonal climate prediction scale. Some of these studies highlighted a need for the means to measure the added value, on the GCM/ESM, by the RCM. One such method used to assess the added value, as recommended by Gnitou et al. (2021), is by calculating the distance between the driving GCM/ESM and the corresponding RCM using any statistic. This method has also been used by Tamoffo et al. (2020) for process-based analysis of the added value making it suitable for application on the ITCZ process. Following Gnitou et al. (2021), we have chosen the KGE, a measure of the model skill, to assess the added value (AV) by CRCM (AV = KGE_{CSIRO ESM} minus KGE_CRCM) in identifying the position of the ITCZ. The added value by CRCM remains positive for all the seasons and up to as high as 4% (during La-Niña NDJ) and 6% (during El-Niño DJF) (Figure 3h). This shows a positive addition of value of up to 6% by CRCM from CSIRO ESM, indicating a great progress in African seasonal climate prediction using a dynamically downscaled ESM. This is even though some authors (Gnitou et al., 2021) would rather prefer a 10% and higher added value.

The Taylor diagram (Figure 3g) compares CHIRPS, CRU and ERA5 datasets to the CRCM predictions over the four rolling seasons (NDJ, DJF, JFM, FMA) to analyze the model performance. The ERA5 dataset was used as the reference dataset and normalizes the standard deviations of CHIRPS, CRU, and CRCM. The model’s fair track of the spatial ITCZ position is confirmed by the correlation coefficients that range from 0.39 to 0.89 at 99% confidence level. Furthermore, the Taylor diagram shows the weakening correlations between the observations/reanalysis with lead time except for CRCM against ERA5 where the correlation weakens into DJF but strengthens again in JFM and FMA. The standard deviations are comparable with Observation datasets showing lower standard deviations than CRCM in NDJ and DJF while ERA5 reanalysis shows lower standard deviation for all seasons except FMA (during El-Niño and La-Niña) wherein CRCM shows a lower standard deviation than ERA5 reanalysis. The RMSE is lower than 1.5° for all seasons and verifying datasets with the lowest RMSE occurring in FMA (CRCM during El-Niño and La-Niña) and JFM (CRCM during El-Niño) where it is less than 0.5°. These values, in comparison to Figure 2e, show that CRCM has a consistent performance between climatology and the ENSO phases. This CRCM performance on the ITCZ position reflects a very significant progress in regional dynamical climate modeling over southern Africa in terms of the onset and offset of seasonal precipitation which is linked to the position of the ITCZ (Randriatsara et al., 2022).

The observations/reanalysis datasets correlations during El-Niño show reasonably better association between CHIRPS and CRU (r ≥ 0.96) than CHIRPS-ERA5 & CRU-ERA5 correlations (0.47–0.81 and 0.48–0.79 respectively). These correlations are comparable to the CHIRPS-CRU and CRU-ERA5 correlations found by Steinkopf and Engelbrecht (2022) for DJF when correlating CHIRPS, CRU, and ERA5 over the tropical (between 20°S and 20°N) Africa.

The spatial spearman correlations between CRCM and individual observation/reanalysis dataset range between the lowest of 0.39 (CRCM and CRU during FMA) and highest of 0.88 (CRCM and ERA5 during FMA). CRCM diverges more than CSIRO ESM, with the lead time, from the observations/reanalysis (Figures 3a–d). This indicates the faster error growth by CRCM, in identifying the spatial ITCZ position, than CSIRO ESM further from the initialization time. With more focus on the flow pattern between CSIRO ESM/CRCM and observations/reanalysis datasets, it is evident that both CSIRO ESM and CRCM perform equally well in tracking the spatial ITCZ (Figures 3a–d) as also evidenced in the comparable correlations from Table 2. This agrees with the findings of Tapiador et al. (2020): it is now possible to run an ESM with a higher resolution and attain the same results as the RCM with an added advantage of avoiding most issues common with the RCMs.

This section presents the analysis of CRCM performance in detecting the regional edge of the HC, the regional STJ, and multi-level streamfunction at different lead times and ENSO phases. These characteristics form part of the teleconnections between the ITCZ (ascending edge of the HC, Faulk et al., 2017; Schwendike et al., 2014), descending edge of the HC and the STJ. Therefore, it is important to evaluate CRCM performance on these as part of the larger exploratory metrics (Pendergrass et al., 2020), especially considering the consistency of performance between the CSIRO ESM and CRCM as well as the skill and reliability of the CRCM on the position of the ITCZ.

The evaluation of CRCM was done with a focus on 10°E–40°E, covering the African tropics (between 20° North and South). This process-based model evaluation seeks to assess the predictive skill and explanatory values which provide a chance to improve process representation by the model (Steyn and Galmarini, 2008).

The differences on the descending edge of the HC latitude between ERA5 and CRCM vary from the highest of 24.3% (NDJ season during La-Niña) to the lowest of 3.5% (FMA season during La-Niña) (Figure 3f). The differences in the STJ latitude between ERA5 and CRCM vary between the highest of 7.5% (JFM season during La-Niña) and the lowest of 0.1% (DJF season during La-Niña). These MAPE values are within the error margin for a highly accurate to a good forecasting model (Montaño Moreno et al., 2013; Zhao et al., 2023). This means CRCM is performing well enough to study the width and area of the ITCZ.

SMAPE was used for the mean meridional zonal wind to eliminate the sensitivity of the results to outliers resulting from the high variability of the differences between the ERA5 reanalysis and CRCM winds prediction over Africa. The response of the model to lead time is visible between 740 hPa and 590 hPa (marked by purple dotted lines in Figure 5h) while no clear response is seen throughout all other levels. The model performance through these other levels is almost similar over all seasons. The CRCM reflects the consistently lower SMAPE from 300 hPa upwards and this result is aligned to the performance of the model in identifying the STJ latitude (Figure 5g), showing a maximum MAPE of 7.5% during JFM season. The changes (less than 3°) in the latitude position of the STJ as investigated here are aligned to the previous findings by Manney et al. (2021) that the STJ latitude changes relating to ENSO are within the 3°, over the southern hemisphere. Furthermore, we found that the STJ latitude shifts poleward during El-Niño while it shifts equatorward during La-Niña over the southern hemisphere and Africa. In comparison, the model performs better in identifying the STJ latitude (Figure 5g) than it does in identifying the edge of the HC (Figure 5f).

The performance of the CRCM’s 500 hPa stream function was found to be out of phase with the ERA5’s 500 hPa stream function and mainly shifted southward even though it can correctly pick the pattern with minimal magnitude differences. The phase corrected presentation (Figure 5e) of the climatological 500 hPa stream function by CRCM (dash dotted lines) against the climatological 500 hPa stream function by ERA5 (solid lines) indicates similar pattern, with different magnitudes, and a minor southward shift. In consideration of this, it follows that CRCM can be used to study the strength of the ITCZ. There is a loss of skill, with lead time, by CRCM for both wind patterns and stream function (Figures 5a–d).

The descending edge of the HC is found to shift to the north of its climatological position during El-Niño while it shifts to the south of its climatological position during La-Niña (Figures 5a–d,f). This applies to both the ERA5 dataset and CRCM output. This HC shift is aligned to the shift of the zonal position of the ITCZ during the same phases of ENSO, especially from CHIRPS, CRU and ERA5 datasets. On the contrary, the latitude of the STJ has been found to shift to the south of its climatological position during El-Niño (for both ERA5 and CRCM output) while it shifts to the north of its climatological position in response to La-Niña (for CRCM while this is only the case for NDJ and JFM seasons for ERA5 dataset), Figures 5a–d,g. The descending edge of the HC (−32.8° by CRCM and −36.5° by ERA5) during DJF season is comparable to that from Mahlobo et al. (2018) (−36°) even though they used the ERA-Interim dataset. The response of the local HC to ENSO has previously been reported by Schwendike et al. (2014) who concluded that local Hadley circulation response to ENSO is much stronger than the local Walker circulation response to ENSO. According to Wolf et al. (2021), the ITCZ is a narrow band in which the northerly and southerly trade winds of Hadley circulation converge. This existing association between the HC structure and the ITCZ (Faulk et al., 2017) explains the intertwined shifting relationship between the edge of the HC and ITCZ in response to the ENSO phases.

Calculation of the mean ITCZ position, from different observations and reanalysis datasets, makes it possible to compare the ascending/descending edge from CRCM to the overall mean latitude for the ascending/descending edge of the HC over Africa, from various observation/reanalysis datasets. Table 3, summary of Figures 5a–d,f, indicates that the southernmost latitude for the ascending edge of the HC is found around 10.1°S (9.9°S) and descends around 36.5°S (32.8°S) during DJF, according to the mean ITCZ (CRCM). Similarly, the northernmost latitude for the ascending edge of the HC is found around 6.8°S (5.6°S) and descends around 32.3°S (29.7°S) during FMA as per the mean ITCZ (CRCM). The ITCZ position is also found within latitudes over which ERA5, and CRCM stream function differences are minimal (Figures 5a–d), making it possible to identify the ITCZ position using CRCM stream function (Figures 6, 7).