<p>Observational uncertainty plays a critical role in shaping the reliability of weather forecasts and climate projections. In this study, we identify the most reliable dataset to reduce observational uncertainties for climate projection based on performance of the Statistical Downscaling Model (SDSM). Calibrated with four gridded precipitation datasets (APHRODITE, AgERA5, IMD, and CHIRPS), spanning the period from 1979 to 2014, the study evaluates both the calibration and validation phases using key evaluation metrics: R², NSE, RMSE, and PBIAS. A novel composite ranking system that integrates multiple evaluation indices was developed, reducing the complexity of multi-model, multi-metric comparisons and enhancing the reliability of model evaluation. 90% of the models calibrated with APHRODITE data demonstrate consistency in estimating precipitation with 57% of these models under Rank I. The results indicate that model calibrated with the APHRODITE and AgERA5 datasets exhibit strong performance in mean rainfall estimation and percent wet days. In contrast, models calibrated with the IMD and CHIRPS datasets require substantial improvement in maintaining variance and simulating extreme rainfall events. For predicting maximum precipitation, SDSM is found to be unsuitable as it lacks the ability to preserve variance in projections. These findings underscore the critical importance of dataset selection in downscaling applications and highlight the need for improved observational data to minimize uncertainties in climate projections.</p> Graphical Abstract <p>The graphical abstract illustrates the study in three integrated stages: first, NCEP-DOE reanalysis predictors are combined with multiple gridded precipitation datasets (APHRODITE, AgERA5, IMD, CHIRPS) within the SDSM framework, with calibration (1979–2004) and simulation (2005–2014) periods, and descriptive statistics—mean, maximum, variance, percentiles, wet days, and wet spell length—used to assess dataset performance. Second, these statistics are evaluated through performance indices (R², NSE, RMSE, PBIAS) against thresholds, with a rating table classifying model skill from Very Good to Unsatisfactory, highlighting a comparative diagnostic approach to identify robust datasets. Finally, a composite classification integrates these indices into five model quality categories, revealing APHRODITE as the most reliable dataset (57.1% in Category I, Very Good), with the stepwise framework enabling pointwise analysis and distinction between uncertain and reliable precipitation data for climate projections.</p> <p></p>