What is the optimal length of the calibration period?

When applying a hydrological model, the length of the calibration period is typically based on the length of the available hydroclimatic data series. Usually, half of the data are used for the calibration of the model and the other half for validation, but other splits (e.g., three quarters and one quarter) are possible as well. When the data record is short, all data may be used for calibration. The general idea is that a longer calibration period will include a wider range of conditions (e.g., a wider range of flood events) and thus lead to a more robust model. However, a longer calibration period does not always have to be better. There are reasons for not using a (too) long calibration period. First, a long calibration period may not be necessary if the extra years of data do not contain any additional information (i.e., different conditions). In this case, a longer calibration period may just waste computer resources, which is an issue when the model is calibrated for a large number of catchments. Second, some discharge records are by now more than 80 years long. During this time period many things have changed. This includes the way that streams are gauged, leading to differences in data accuracy. The catchments themselves will likely have changed as well. For some catchments, these changes are obvious but for other catchments they are more subtle. Even if the dominant land use has remained agriculture, the agricultural practices have changed. Similarly, for catchments that have remained forested during the period of data collection, there may be changes in the percent or spatial pattern of open areas or changes in the species composition. One could, therefore, argue that there is a trade-off between a long calibration period that includes all the variation in the climate and not using data from a period during which the catchment was different from the current conditions. With increasing length of available data series the question on the optimal length of the calibration period becomes more relevant.

To explore the sensitivity of the model results to the length of the calibration period, we calibrated the HBV model for several Japanese and Swiss catchments for which long hydroclimatic records are available. We split these records into multiple calibration and validation periods of different lengths and assessed 1) how the drop in model performance between the calibration and validation period depends on the periods chosen for model calibration and validation, and 2) how the length of the calibration period affects the range in model calibration and validation performances. The results show that the optimal length of the calibration period depends on the catchment, and differs even for neighboring catchments. These analyses provide some information on the optimal length of the calibration period for the study catchments but need to be repeated for other catchments to prove the generalizability of the results.