Hi Saima,

For validation, you may check for RMSE, correlation coefficient with observed vs simulated data. Also it may be recommended to carry out a homoskedasticity assessment for the residuals (predicted vs residuals) to see if they form a trend or not (if there is a trend in your error then there is a problem). Likewise, you can check the normality of your residuals to see if they are random (naturally occurring residuals are random) or biased.

I have one more suggestion, instead of manually dividing the calibration and validation data, I think it makes more to make a random statistical sampling and select the calibration and validation data randomly. Say,, you have 4000 observations, randomly sort them and then say, use 2500 observations to calibrate the model and the remaining 1500 to validate it. This way you are sure that your selection itself is not biased. If you select the calibration/validation data based on time or any other parameter, you'd expect your model to be biased. This is a general suggestion I am not sure about your actual case specific issues and concerns.

Thanks!