Weak seasonality perspective on blood collection and supply demand forecasting methodology
10.13303/j.cjbt.issn.1004-549x.2026.08.011
- VernacularTitle:弱季节性视角下的采供血需求预测方法论
- Author:
Wujin SU
1
;
Nan CHEN
1
Author Information
1. Nanning Blood Center, Nanning 530007, China
- Publication Type:Journal Article
- Keywords:
model competition;
demand forecasting;
weak seasonality;
methodology;
ARIMAX;
DTW clustering;
attribution paradigm
- From:
Chinese Journal of Blood Transfusion
2026;39(8):1067-1073
- CountryChina
- Language:Chinese
-
Abstract:
Objective: To develop a forecasting framework suitable for weak-seasonality time series using the classical autoregressive integrated moving average (ARIMA) model as the baseline. Rather than focusing exclusively on predictive accuracy, we propose an interpretive, attribution-oriented approach to blood demand forecasting that emphasizes explanatory power over pure precision. The aim is to provide a domain-specific modeling reference for the blood collection and supply sector. Methods: Monthly clinical red blood cell (RBC) demand data from Nanning, China (January 2018-December 2025) were collected. Seasonal strength was assessed by STL decomposition; stationarity was tested using ADF and KPSS procedures. ARIMA, seasonal ARIMA (SARIMA), Holt-Winters, ARIMAX, and a hybrid ARIMA-XGBoost model were built and compared in terms of model structure, forecast error, and operational interpretability. Annual monthly series were further subjected to DTW-based K-means morphological clustering to examine inter-annual variation under weak seasonality. Results: SARIMA auto-selection resulted in an ARIMA (4, 1, 0) specification. The Holt-Winters seasonal smoothing parameter γ was 1.00, and the seasonal strength index (F
) was 0.18, confirming weak seasonality. DTW-KMeans clustering (k=3) identified three phases: a low-demand cluster (2020-2022, annual mean≈27 002.3 U), a medium-demand cluster (2018 and 2023, annual mean≈38 755 U), and a high-demand cluster (2019 and 2024-2025, annual mean≈44 284 U), suggesting that weak seasonality mainly reflects shifting clinical demand patterns across different stages. In out-of-sample validation, ARIMAX achieved the best performance (RMSE=8 374.88; MAE=6 669.88; MAPE=14.31%), significantly outperforming ARIMA (Diebold-Mariano test P=0.011). Time series cross-validation further confirmed that ARIMAX performed best in the last two folds, with the smallest standard deviations in RMSE (8, 954.46±3, 530.61) and MAPE (18.75%±6.83%), indicating superior predictive stability. The ARIMA-XGBoost hybrid showed no meaningful improvement (RMSE=10, 802.20; MAPE=21.99%) and performed worse than standalone ARIMA. Conclusion: For weakly seasonal blood demand series with structural change, ARIMAX—which explicitly incorporates explanatory variables—offers better stability and practical relevance than purely data-driven or hybrid machine learning alternatives. We recommend that blood demand forecasts test seasonal assumptions before model selection, encode domain knowledge as exogenous inputs where possible, and critically assess whether machine learning hybrids are warranted in low-seasonality settings.
