Performance evaluation of bias correction methods for CMIP6 precipitation datasets in the drylands of the Wolaita Zone, Southern Ethiopia

The Coupled Model Intercomparison Project Phase 6 (CMIP6) provides advanced general circulation models (GCMs) to model interactions among climate systems under mathematical formulations (Johnson and Sharma, 2011; Lanzante et al., 2018). These models are used to model future climate conditions and assess their impacts on ecosystems, water resources, agriculture and human health (Alizadeh-Choobari, 2019; Trzaska and Schnarr, 2023). For instance, recent studies have used CMIP6 outputs to contrast future ecosystem productivity under various shared socioeconomic pathway (SSP) scenarios (Qiu et al., 2023).

Nonetheless, GCMs are characterized by their coarse spatial resolution, typically 80–250 km (Cannistra, 2023). This limitation restricts their depiction of regional processes, such as orography, land–sea contrast and convective cloud systems, which are of the utmost priority for regional applications, such as hydrology and agriculture planning (Alizadeh, 2022; Gutmann et al., 2014). To ease this, high-resolution setups, such as the HadGEM3 GC31 HM with a resolution of approximately 50 km, have been developed, and downscaling via machine-learning techniques has been applied to achieve finer spatial resolutions of approximately 0.1° (Wang et al., 2025).

However, GCM projections are often marked by random and systematic biases that, if not corrected, can lead to misleading assessments and ineffective adaptation strategies (Loikith et al., 2023; Aznar-Siguan et al., 2024). Therefore, downscaling and bias correction methods (BCMs) are required to increase the credibility of climate projections. Examples of ordinary approaches include statistical approaches such as scaling approaches, quantile mapping and dynamical downscaling (Chen et al., 2025). The method of choice relies on the intended application, data availability, computer ability and target climate variable (Tapiador et al., 2020).

Although CMIP6 represents a remarkable improvement over the older CMIP3 and CMIP5 models (Gusain et al., 2020; Wang et al., 2024), there are still uncertainties in the tropical and subtropical regions (Kharin et al., 2013). Recent evaluations indicate that uncorrected CMIP6 precipitation data should not be employed directly without bias correction (Nguyen-Duy et al., 2023). Some BCMs have been tested globally, including local intensity scaling (LIS), multiplicative linear scaling (MLS), distribution mapping (DM), multiplicative delta change (MDC) and power transformation (PT) (Luo et al., 2018; Tefera et al., 2023). They behave differently depending on the place, data set and method (Lafferty et al., 2021; Takhellambam et al., 2023), implying that there is no single approach.

Various BCMs have been used to regional and global climate model outputs in Ethiopia, but with mixed outcomes (Daniel, 2023; Enyew et al., 2024). For example, Daniel (2023) conducted an experiment with MLS, PT and DM for southern watersheds, whereas Enyew et al. (2024) applied PT to CMIP6 data for the Upper Blue Nile. However, few studies have compared different BCMs in Ethiopia’s drylands, where livelihoods largely depend on climate variability. This study aims to fill this gap by comparing five BCMs, namely, MLS, MDC, LIS, PT and DM, applied to CMIP6 precipitation data for the Wolaita Zone Drylands. By observing what improves or worsens model performance, this study can be used to support better regional climate studies and sustainable development objectives, particularly climate action.

This study was conducted in the drylands of the Wolaita Zone, southern Ethiopia, focusing on four districts: Duguna Fango, Kindo Koysha, Humbo and Ofa (Figure 1). The analysis was based on data from four meteorological stations located within these districts.

The Duguna Fango district is situated on the eastern side of Wolaita, facing the Bilate River and the Central Rift Valley, which strongly influences its hydroclimate and land-use dynamics. The district seat, Bitena, is located approximately 42 km east of Wolaita Sodo and 300 km south of Addis Ababa by road. The Bilate meteorological station (6.822° N, 38.088° E; 1,361 m a.s.l.) is located within the Duguna Fango district, approximately 251 km south-southwest of Addis Ababa and 36 km east of Wolaita Sodo. The Kindo Koysha district borders the Dawro and Koysha districts to the north and east, respectively, and spans elevations around 1,500 m a.s.l. The district’s administrative seat is Bele (Bale Hawassa). The Bele meteorological station (6.918° N, 37.526° E; 1,240 m a.s.l.) is located within the Kindo Koysha district, approximately 266 km south-southwest of Addis Ababa and 27 km west-northwest of the district’s administrative seat, Bele (Bale Hawassa).

The Humbo district stretches across Rift Valley lowlands and mid-altitude uplands, offering diverse ecological conditions. The district seat, Tebela, is located approximately 17 km south of Wolaita Sodo and hosts the Humbo station (6.703° N, 37.766° E; 1,618 m a.s.l.), which is situated within the Humbo district and approximately 276 km south-southwest of Addis Ababa. The Ofa district, located on the southwestern flank of Wolaita, shares borders with the Kindo Koysha and Kindo Didaye districts. Its administrative seat is Gessuba, which is also home to the Gessuba station. The coordinates of the Gessuba station are 6.729° N, 37.563° E, with an elevation of 1,552 meters above sea level. It is situated 26 kilometers west-southwest of Wolaita Sodo and 283 kilometers south-southwest of Addis Ababa.

The Wolaita Zone drylands experience three climatic seasons: winter – Bega (October–January); spring – Belg (February–May); and summer – Kiremt (June–September) (NMI, 2023). The zone has extensive sections of spring and summer rainfall – 82% of the annual total. Winter is a relatively dry season with the lowest contribution to the rainfall for the year – 18%. Interannual and seasonal rainfall over the 1990–2021 interval suggests moderate rainfall variability (CV 10%–20%), whereas relatively greater variability in the winter season (CV > 30%) is felt. Low interannual and seasonal variability in study area temperature has been felt over the same interval. A daily maximum (day) temperature ranges from 24.3°C in July to 30.1°C in February, and a minimum (night) temperature ranges from 12.6°C in December to 15.1°C in April (Figure 2).

Daily precipitation data from four meteorological stations spanning the years 1990–2021 were obtained from the Ethiopian Meteorological Institute. To ensure the accuracy and reliability of the data, a rigorous quality control process was implemented. Missing values were filled using multiple linear regression, a commonly used method in Ethiopian climate research to maintain consistency across both space and time (Chinasho et al., 2021). Outliers were identified using the interquartile range method, which effectively distinguishes erroneous entries without excluding genuine extreme values. The homogeneity of the data sets was then verified using the Pettitt test (Pettitt, 1979) and Buishand’s range test (Buishand, 1982). These non-parametric tests are well-established for detecting shifts that may be caused by station relocation, equipment changes, or other non-climatic factors. By using both tests, the author was able to reliably detect any inconsistencies in the data. This multistep procedure resulted in a consistent and high-quality data set suitable for long-term variability analysis.

Daily precipitation data from six CMIP6 GCMs (EC Earth3 Veg, ACCESS CM2, ACCESS ESM1-5, HadGEM3 GC31-LL, MIROC ES2L and MPI ESM1-2 HR) were utilized for both historical (1990–2014) and future (2025–2100) periods, under two scenarios: SSP2-4.5 (medium stabilization) and SSP5-8.5 (high emission). These models were specifically chosen due to their proven ability to accurately simulate rainfall patterns in Ethiopia and East Africa, as demonstrated in previous evaluations (Rettie et al., 2023; Lebeza et al., 2024; Gashaw et al., 2024). Their inclusion in this study enhances representativeness and reduces uncertainty related to model selection.

The climate model data was processed using CMhyd software, which combines station observations with GCM outputs to downscale and correct for biases. CMhyd has become a popular tool in Ethiopian hydrological and climate studies due to its effectiveness in preserving local climate patterns (Tullu and Tumsa, 2024; Lukas et al., 2025). For this study, data was extracted from six models, two scenarios and five bias correction methods, resulting in a total of 60 corrected data sets.

Five commonly used BCMs were evaluated: MLS, MDC, LIS, PT and DM. These methods were chosen because they represent a range of correction approaches, from simple mean-based methods (MLS, MDC) to techniques that focus on the distribution of data (DM, PT). LIS is particularly useful for adjusting daily precipitation intensity and frequency, which is crucial for analyzing extreme events. The performance of each method was assessed using three metrics: coefficient of determination (R²), Nash–Sutcliffe efficiency (NSE) and mean absolute error (MAE). These metrics collectively measure the strength of correlation, predictive accuracy and average deviation, providing a comprehensive evaluation of the model’s skill. The use of both variance-sensitive and error-sensitive indicators enhances the reliability of the bias correction assessment.

To analyze extreme rainfall, the 10th and 90th percentiles of daily rainfall were calculated using a non-parametric thresholding approach. This method is commonly used in climate research, as it does not rely on assumptions about the underlying distribution, making it well-suited for regions with high variability in rainfall (Dinku et al., 2024; Mengistu et al., 2024). By aggregating these values on a monthly basis, a meaningful comparison between observed and bias-corrected data sets was possible.

The accuracy of CMIP6 precipitation outputs varied considerably across the selected models, as shown in Table 1. The application of the MLS method significantly improved rainfall simulation performance for several individual models, as well as for the ensemble mean. Specifically, there were notable improvements in key evaluation metrics, including the NSE and the coefficient of determination (R²). For example, MLS improved ACCESS ESM1 5 by 55.92%, EC Earth3 Veg by 10.39%, MIROC ES2L by 76.98%, MPI ESM1 2 HR by 1.38%, and the ensemble mean by 41.07%. These results demonstrate the effectiveness of MLS in reducing systematic biases in CMIP6 precipitation outputs while maintaining the temporal structure of rainfall.

These findings are consistent with previous studies. For instance, Amognehegn et al. (2023) assessed the impact of MLS bias correction on daily precipitation from twelve CMIP6 models and reported significant accuracy gains in three models, including ACCESS ESM1 5. Similarly, Chae and Chung (2024) evaluated twenty CMIP6 models across the Mokgam River watershed and found that MLS substantially improved the fidelity of simulated precipitation for models such as ACCESS CM2, ACCESS ESM1 5, EC Earth3 Veg and MPI ESM1 2 HR. The alignment between the present results and existing literature reinforces the robustness of MLS as a bias correction method, particularly in regions with highly variable rainfall.

While the mean precipitation values improved with the use of MLS, it was not as effective in correcting extreme precipitation indices. This can be seen in Figure 3, where MLS had limited capability in improving the accuracy of the 10th and 90th percentiles. This is due to the linear, mean-based nature of MLS, which does not account for higher-order distributional properties such as variance or skewness. This is supported by a study by Azad and Ahmadi (2024), which found that MLS performed poorly in simulating precipitation extremes across 13 global climate models in Iran, including ACCESS-ESM1-5 and EC-Earth3-Veg. In their research, MLS increased the MAE by 3.75% for ACCESS-CM2 and by a significant 35.78% for HadGEM3-GC31-LL, highlighting the limitations of the method for extreme events. Despite these limitations in individual models, the ensemble mean showed more robust results.

After bias correction, the ensemble outputs showed notable reductions in MAE compared to the six individual models: −24.7% (ACCESS-CM2), −22.5% (ACCESS-ESM1-5),−15.1% (EC-Earth3-Veg), −49.2% (HadGEM3-GC31-LL), −9.9% (MIROC-ES2L) and −7.8% (MPI-ESM1–2-HR). This is consistent with previous research that has shown that ensembles better capture both central tendencies and variability (Lafferty et al., 2021; Takhellambam et al., 2023). Overall, these results demonstrate both the strengths and limitations of MLS. While it significantly improves mean precipitation simulation, its linear approach limits its performance for extreme rainfall events. These limitations highlight the importance of considering more advanced or non-linear bias correction techniques when modeling precipitation percentiles, which are crucial for drought and flood risk assessment. Furthermore, the superior performance of the ensemble mean suggests that combining multiple models enhances reliability and reduces model-specific biases, emphasizing the value of ensemble approaches in climate impact assessments and water resource planning in regions like southern Ethiopia.

The MDC method significantly improved the accuracy of CMIP6 precipitation outputs for daily rainfall in southern Ethiopia, as shown in Table 2. Prior to bias correction, CMIP6 models performed poorly, with MAE ranging from 169.55 to 1321.66 mm, coefficients of determination (R²) between 0.005 and 0.17, and NSE values spanning from −72.88 to −0.626. However, after applying MDC, these metrics improved significantly: MAE decreased to 9.61–96.82 mm, R² increased to 0.99 and NSE rose to 0.71–0.99, indicating a strong alignment with observed data. The error reductions varied among models, with the ensemble mean showing the highest improvement of 93.7%, followed by ACCESS-CM2 with 49.5%. The ensemble generally outperformed individual models, except for MPI-ESM1-2-HR. These findings are consistent with previous regional studies. For example, Babiker et al. (2024) reported error reductions of 0–18% after applying MDC in Sudan, demonstrating the method’s applicability in the region.

Similarly, Tsegaye et al. (2023) found that more than 70% of CMIP6 models showed NSE values above 0.80 and R² above 0.90 after MDC was applied in Ethiopia’s Upper Awash Basin. Ali et al. (2025) further confirmed MDC’s consistent ability to reduce seasonal rainfall biases in East Africa, particularly during the Belg and Kiremt seasons. Rauf and Desta (2023) highlighted MDC’s suitability for arid and semi-arid zones like southern Ethiopia, where rainfall variability and nonlinearity limit the effectiveness of simpler correction methods. These studies collectively support the statistical robustness and adaptability of MDC for regional climate modeling applications.

Although the MDC effectively corrected the mean precipitation, its performance varied across precipitation extremes. This method had a limited impact on the simulation of the 10th percentile, representing dry conditions, across the six CMIP6 models (Figure 4). This limitation is likely due to the linear approach of the method, which does not fully account for skewness or variability at the lower end of the precipitation distribution. Conversely, the MDC performed exceptionally well for the 90th percentile, representing extreme wet conditions, achieving an MAE of 0.02 mm, an R² of 0.9 and an NSE of 0.9. This demonstrates the MDC’s strength in adjusting high-end precipitation estimates, which is consistent with the findings of Oruc (2022), who reported the superior performance of the MDC over quantile mapping (QM) and empirical quantile mapping (EQM) in Turkey.

Teklemariam et al. (2023) found that MDC outperformed other models in bias-correcting high precipitation quantiles over the Rift Valley Basin, including southern Ethiopia. Desta et al. (2024) reported that MDC preserved interannual variability while minimizing overcorrections under high-emission scenarios. Similarly, Feyissa and Lemma (2025) demonstrated enhanced performance in percentile-based skill scores when applying MDC to CMIP6 data sets in semi-arid Ethiopian zones. Together, these studies highlight MDC’s growing recognition of the MDC as a dependable method for improving high-intensity rainfall simulations in regional climate applications.

In summary, this method provides substantial improvements in mean precipitation simulation and effectively corrects high-end extremes, although it has limited success in correcting low-percentile rainfall. The ensemble mean further enhances reliability by combining multiple models, mitigating individual model biases, and providing a more robust representation of rainfall variability.

The LIS bias correction technique produced mixed outcomes when applied to CMIP6 precipitation data sets for daily rainfall forecasting in southern Ethiopia. Specifically, LIS substantially reduced error rates by 50.7%, 7.1%, 77.1% and 39.4% for the ACCESS-ESM1-5, EC-Earth3-Veg, MIROC-ES2L models and the Ensemble, respectively (Table 3). However, LIS increased the error rates by 5.9% for ACCESS-CM2, 38.6% for HadGEM3-GC31-LL and 1.6% for MPI-ESM1-2-HR, highlighting the model-specific performance limitations. These findings are consistent with those of earlier studies by Addor et al. (2014), Lafferty et al. (2021) and Takhellambam et al. (2023), who demonstrated that bias correction techniques can sometimes inadvertently amplify uncertainties or introduce new biases.

Further evidence from regional studies supports these results. Belay et al. (2023) reported that LIS effectively reduced biases in high-flow rainfall but worsened errors in dry-day frequency simulations in the Abaya-Chamo Basin. Similarly, Tesfaye and Endeshaw (2024) found that LIS improved the correction of high-intensity rainfall events in some CMIP6 models but performed poorly for lower quantiles, leading to overestimated dry spells. Mohammed et al. (2025) also concluded that while LIS enhanced precipitation accuracy in unimodal rainfall regions, it could overcorrect and distort rainfall variability in bimodal systems such as those in southern Ethiopia. Collectively, these studies underscore the importance of region-specific calibration and model-dependent application of bias correction methods such as LIS.

When examining precipitation extremes, the LIS technique enabled the CMIP6 GCMs to effectively simulate the 10th percentile of observed rainfall in southern Ethiopia (Figure 5), demonstrating improved performance in capturing low-end precipitation extremes. The LIS consistently performed well across individual GCMs and the ensemble mean, highlighting its robustness in correcting light rainfall and reducing dry-day overestimations common in uncorrected CMIP6 outputs for arid and semi-arid regions.

In contrast, LIS failed to improve the accuracy of the 90th percentile of monthly rainfall, particularly in the dry areas of the Wolaita Zone. This limitation reflects the difficulty in adjusting upper-end extremes, consistent with Takhellambam et al. (2023), who reported that LIS struggles with nonlinear biases in heavy rainfall regimes. Supporting evidence from recent studies shows a similar: Hailemariam et al. (2023) noted that while LIS accurately corrected the 10th percentile rainfall in the southwestern Ethiopian Highlands, it underestimated extreme events.

Demelash and Hassen (2024) found that LIS was effective for low to moderate rainfall but inadequate for high quantiles in southern Ethiopia’s bimodal systems. Likewise, Gizachew et al. (2025) emphasized the LIS’s limited adaptability to non-stationary biases during high-rainfall months, restricting its usefulness in flood-risk modeling and extreme weather forecasting in dryland areas such as Wolaita.

Overall, these results indicate that LIS is most suitable for correcting lower precipitation quantiles, particularly in arid and semi-arid climates, whereas alternative or hybrid bias correction methods may be necessary to address high-end precipitation extremes under changing climate conditions.

The PT technique exhibited varied impacts on the performance of CMIP6 models in bias-correcting precipitation data across southern Ethiopia. While PT enhanced the performance of certain models, it led to increased error rates in others (Table 4). Specifically, the error rates increased by 5.6%, 35.1% and 3.8% for ACCESS-CM2, HadGEM3-GC31-LL and MPI-ESM1-2-HR, respectively. In contrast, PT significantly reduced errors by 46.5%, 0.3%, 73.6% and 35.8% for ACCESS-ESM1-5, MIROC-ES2L and the ensemble average, demonstrating a model-dependent efficacy. Supporting this, Enyew et al. (2024) highlighted that PT improved the statistical properties of daily precipitation series, particularly for the MPI-ESM1-2-HR model in the Upper Blue Nile Region. The study demonstrated that PT effectively adjusted skewed rainfall distributions, making it useful for improving projections in both wet and transitional climates in the region.

Similarly, Kebede and Asfaw (2023) reported that PT improved the normality and variance structure of CMIP6 model outputs in central and southern Ethiopia, although they noted that PT tended to over-smooth extreme precipitation events in some models, particularly HadGEM3-GC31-LL. Similarly, Tsegaye et al. (2025) observed that PT enhanced monthly rainfall predictability for models such as MIROC-ES2L but did not fully correct the biases during seasonal transition periods, particularly in bimodal rainfall systems. Moreover, Hailu and Mersha (2025) emphasized that PT performance improves when combined with quantile mapping, suggesting that hybrid correction approaches may be more suitable for handling precipitation extremes in climatically complex regions.

Collectively, these findings indicate that while PT can improve the performance of some GCMs, its application requires careful calibration and validation, particularly in regions with high rainfall variability and nonlinear precipitation distributions.

Despite these improvements in certain metrics, the low R² values and negative NSE scores indicate that PT did not substantially enhance the overall accuracy of any of the six CMIP6 models used in this study. This underscores its limited effectiveness in correcting precipitation data in areas with complex rainfall patterns. PT performed particularly poorly in reproducing the 10th and 90th percentiles of rainfall (Figure 6), suggesting that its assumptions about data distribution transformations are not well suited to the nonlinear, skewed, and seasonally bimodal rainfall typical of southern Ethiopia.

While Chae and Chung (2024) reported substantial improvements using PT for 20 CMIP6 models in South Korea’s Mokgam River watershed, this success appears region-specific, likely due to the more stable, unimodal rainfall regime in that area. Similarly, Beyene et al. (2023) found PT ineffective at correcting both high and low rainfall quantiles in the semi-arid Rift Valley, and Getahun and Wolde (2025) observed minimal improvements in lowland southern Ethiopia. Dinku et al. (2025) further emphasized that PT alone is insufficient and performs better when paired with nonlinear correction methods. Collectively, these results suggest that PT is not recommended for bias correction of CMIP6 models in the drylands of the Wolaita Zone.

Based on the R² and NSE metrics, the DM bias correction method did not substantially improve the accuracy of the CMIP6 outputs for precipitation estimation in the drylands of the Wolaita Zone (Table 5). DM increased the MAE errors for ACCESS-CM2, HadGEM3-GC31-LL and MPI-ESM1 - 2-HR by 7.0%, 37.1% and 9.7%, respectively, indicating that DM is poorly suited for southern Ethiopia’s semi-arid, bimodal climate with high seasonal variability. However, DM improved the performance of ACCESS-ESM1-5 (44.9%), EC-Earth3-Veg (1.4%), MIROC-ES2L (73.8%) and the ensemble mean (30.5%), suggesting that its effectiveness is model-dependent. These findings align with Acharki et al. (2023), who reported that DM enhanced CMIP6 simulations in semi-arid northwestern Morocco, and Li and Li (2025), who observed improved correlation in fourteen GCMs across Canadian climate zones.

Nonetheless, Mekuria and Tadesse (2024) noted that DM performed better in Ethiopia’s unimodal highlands but poorly in lowland drylands because it could not preserve skewed rainfall distributions and seasonal asymmetries. Similarly, Abate and Workineh (2025) reported that the DM failed to correct rainfall extremes in southern and southeastern Ethiopia, whereas Negash et al. (2023) found that unscaled DM led to dry-season overcorrections, distorting drought simulations in pastoral zones. Overall, these results indicate that while DM can be effective in specific contexts, its inconsistent performance – underscores the need for region-specific or hybrid bias correction methods in the drylands.

A key issue arises from the monthly 10th percentiles of observed rainfall being zero, which caused the DM to reproduce zero values across all six CMIP6 outputs (Figure 7, panels B and D). This outcome reflects the method’s inability to differentiate between true dry conditions and model bias, particularly in semi-arid regions with frequent zero-rainfall observations. The poor performance of DM is further evidenced by the performance metrics, where R² was 0, and NSE could not be calculated because of zero denominators in the variance component. In addition, DM increased error levels in the ensemble mean when estimating the 10th percentile of observed rainfall, directly limiting its utility in modeling low-end extremes critical for drought risk assessment and water resource management in arid zones.

As shown in Figure 7, DM also failed to improve the accuracy when replicating the 90th percentile of the observed rainfall across the models. This limitation suggests that DM’s fixed-scaling approach is not sufficiently flexible to adjust for nonlinear biases in precipitation extremes, particularly under variable climatic conditions. Recent research supports these concerns: Tekle and Dagne (2024) found that DM failed to preserve interannual rainfall variability in the southern Rift Valley, leading to distorted seasonal transitions. Similarly, Alemu et al. (2023) reported that DM could overfit historical dry periods, causing under-predictions during emerging wet years in lowland Ethiopia. Woldemariam and Gebeyehu (2025) also demonstrated that DM introduced statistical artifacts in both lower and upper rainfall quantiles when applied to CMIP6 bias correction in the Horn of Africa.

Overall, these findings reinforce that DM is not well suited for use with the selected CMIP6 models in Southern Ethiopia. Its inconsistent performance for both low-end and high-end precipitation highlights the importance of considering alternative or hybrid bias-correction approaches for accurate climate impact assessments in the drylands of Wolaita Zone.

In this study, the performance of five schemes of BCMs of precipitation data was analyzed in detail. The BCMs used were MLS, DM, MDC, LIS and PT. BCMs were applied to the precipitation data of six models from the set CMIP6: MIROC-ES2L, ACCESS-CM2, ACCESS-ESM1-5, EC-Earth3, HadGEM3-GC31-LL and MPI-ESM1-2-HR. The impacts varied by the respective CMIP6 model and evaluation measure used. Of particular interest, several BCMs oddly escalated errors in some models, reinforcing the relevance of method choice for regional climate modeling. Among the five BCMs tested, the MDC method consistently provided superior corrections across all evaluation criteria, followed closely by the ensemble mean approach. MLS and LIS demonstrated specific strengths but were less effective for extreme events. PT and DM were the least reliable for the Wolaita drylands. These findings highlight MDC and ensemble approaches as reliable tools for enhancing the credibility of CMIP6 precipitation projections in southern Ethiopia. However, it must be remembered that biases arising from insufficient modeling still prevail and cannot be entirely eliminated by BCMs.

The adoption of MDC-corrected ensemble projections offers significant benefits for Ethiopia’s climate strategies. Improved rainfall simulations can strengthen the National Adaptation Plan and Climate-Resilient Green Economy strategy by supporting early warning systems, drought preparedness and integrated water resource management. Agricultural planning, irrigation and hydropower development will benefit from reduced uncertainty in rainfall–runoff modeling. Moreover, reliable projections can inform climate-smart investments, insurance schemes and financial instruments. Residual uncertainties in GCMs and emission scenarios remain, but these should be viewed as a call for adaptive policy frameworks rather than a limitation. MDC-corrected ensembles provide a practical decision-support tool that enhances resilience planning and resource management in Ethiopia’s climate-sensitive.

This research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.