For reasons of simplicity, the most commonly used hydrological models are based on the Soil Conservation Service Curve Number (SCS-CN) model, which is probably a good choice for the estimation of runoff on the Loess Plateau of China; however, the high spatial heterogeneity, mainly caused by a fragmented landform and variations in soil type, may limit its applicability to this region. Therefore, applicability of the SCS-CN model to a small watershed, Liudaogou on the plateau, was evaluated and the most appropriate initial abstraction ratio (I~/S) value in the model was quantified by the inverse method. The results showed that the standard SCS-CN model was applicable to the estimation of runoff in the Liudaogou watershed and the model performance was acceptable according to the values of relative error and Nash-Sutcliffe efficiency. The most appropriate Ia/S value for the watershed was 0.22 because with this modified Ia/S value, the model performance was slightly improved. The model performance was not sensitive to the modification of the Ia/S value when one heavy rainfall event (50.1 mm) was not considered, which implied that the model, using a standard Ia/S value, can be recommended for the Liudaogou watershed because single rainfall events exceeding 50 mm seldom occurred in that region. The runoff amount predicted for the Liudaogou watershed by the SCS-CN model, using the modified Ia/S value, increased gradually with increasing rainfall when rainfall values were lower than 50 mm, whereas the predicted amount increased rapidly when the rainfall exceeded 50 mm. These findings may be helpful in solving the problem of serious soil and water loss on the Loess Plateau of China.