Asosiy kontentga oʻtish
Maqola

Three ways forecast skill is overstated in small-sample streamflow modelling: Evidence from a data-scarce Central Asian basin

Rustamjon NasridinovDepartment of Digital Educational Technologies, Tashkent University of Information Technologies named after Muhammad al-Khwarizmi, Amir Temur Street 108, Tashkent 100200, Uzbekistan
2026en
Ochiq kirishOltin

Litsenziya: cc-by-nc-nd

ABI

Annotatsiya

Reported skill in data-driven streamflow forecasting is often assembled from choices that are individually reasonable but jointly generous. We quantify three of them using 56 years of 10-day (decadal) discharge for the upper Zarafshan, Central Asia, with ERA5-Land forcing aggregated to the same step. First, variational mode decomposition (VMD) applied to the complete series before partitioning, an arrangement common in hybrid studies, raises the Nash–Sutcliffe efficiency (NSE) of a VMD-based network from 0.887–0.908 to 0.988 and cuts its RMSE by a factor of about three. Implemented causally, the hybrid falls below every model that does not decompose. Second, the naive floor is far higher than usually acknowledged: 10-day seasonal climatology alone reaches NSE 0.871 at this resolution, so the best model's 0.944 amounts to a 57% reduction in error variance relative to climatology, not the near-perfect skill the raw number suggests. Third, the seven architectures we compared, from random forest to an attention-based CNN-BiLSTM, span 0.919–0.944 in NSE, and no pairwise Diebold–Mariano test survives Holm correction. For some pairs the uncorrected verdict depends on which random seed is compared. A plain GRU beats both reference predictors significantly, and it does so within each gauge record separately. In this basin, at this horizon and under this protocol, the added complexity does not buy measurable skill.

Hali tarjima qilinmagan

Identifikatorlar

Iqtiboslar va manbalar