Hi Rajesh, since you are using a statistical method, it is, in principle, only a matter of choice and rationale how many predictor variables you use. Being statistical you are not expecting a physical description of the process (in this case the downscaling) and although it is desirable to have a relations between predictor and response variables consistent with theory and the physics of your system, it is not a most, it is a matter of choice. 

That being said, you should try to use as few variables as you need to describe your system/process as accurately as you desire. Depending on your final goal, you may also want to have some theoretical backup for whatever relations between predictor and response variables you have found and finally be carefull to avoid overfitting your data. 

In all statistical and even black-box models, one of the significant points is to have a set of predictors (estimators) which has strong correlation (linear or nonlinear) with dependent variable(s); besides, the selection of predictors (estimators) should be able to justify the physical concepts of the problem. I think regardless of specific number of predictors, you should choose the most appropriate variables in terms of physics as well as model accuracy.

Thank you so much Dr. Rafael and Dr. Mohammad. 

SDSM is not a black-box model, it is a decision based model. Selection of predictors only on the basis of correlation some times won't give satisfactory out-put. The local knowledge is also invaluable when determining sensible combination of predictors. For ex: if you won't get good correlation for mean sea level while you are downscaling precipitation still you have to take mean sea level pressure as a predictors as it effects the precipitation. You can use recently updated SDSM-Decision Centric ( March 9th 2015) package for downscaling  and stress testing.

I strongly agree with Jeew. For some of the predictors which have strong influence on occurrence of precipitation, you have to include them even if they do not have good correlation. But your knowledge of the area is critical as rain bearing systems vary across different geographic areas. 

Are you using SDSM as it is or have your tried the MATLAB Addin (ASD: Automatic Downscaling tool). It makes things a bit easier.  

You can select the predictors as many as possible but please check the correlation and P-value. If you are working with Precipitation please make sure if you choose the process "conditional".