Calibration and validation mismatch

[Anna Zam kottayil]

I am currently working on SWAT model calibration and validation using SWAT-CUP, and I would like to seek your guidance regarding an issue I am facing with the model performance.

For my study, I have used data from 2015 to 2025, considering 2015 as the warm-up period. During calibration, I obtained a Nash–Sutcliffe Efficiency (NSE) value of 0.82, which indicates a very good model performance. However, during validation, the NSE value dropped to 0.51.

I understand that a decrease in NSE during validation is expected, but I am concerned about whether this level of reduction is acceptable or if it indicates potential issues in my model setup. I would like to know if this discrepancy could be due to factors such as parameter overfitting during calibration, variability in input data, or issues related to the selected calibration/validation periods.

Could you please guide me on:

- Whether this validation performance is acceptable for hydrological modeling studies

- Possible reasons for the drop in NSE during validation

- Any steps I can take to improve the validation results

I would really appreciate your insights and suggestions.

Thank you for your time and support.

[omc...@gmail.com]

Hi,

A brief response to your concerns:

- Yeah, it is very concerning this drop, even though that NSE of 0.51 could be label as satisfactory/decent, dropping from 0.82 during calibration will rise some eyebrows.

- I can suggest these possible reasons: overfitting in calibration; calibration and validation stats do not align (imagine that calibration was doing during flood, and validation during drought); and/or one hydrological process is not being consider (i.e. evapotranspiration).
- I will suggest you to check the cal/val stats properties to amke sure that you are trainig and validating your model with the same process (you can define your periods with gaps, which is ok, something like that: 2014-2016 CAL; 2017 VAL; 2018-2021 CAL; 2022-2024 VAL; 2025 CAL); it is not necessary that calibration and validation to be continuos

Regards,

Oscar M Cabezas-Nivin