A dual-stream multitask deep learning model for supply chain demand forecasting with statistically derived early-warning signals

Accurate demand forecasting is crucial in supply chain management (SCM) as global supply networks become increasingly complex and exposed to disruptions. Traditional models that rely mainly on historical sales may miss demand shifts driven by heterogeneous auxiliary signals and often do not provide explicit early warning of abnormal events. This study proposes a dual-stream multitask deep-learning model that integrates demand forecasting and statistically derived earlywarning signal estimation within a unified framework. The model adopts a dual-input architecture: (i) an LSTM branch that captures temporal dynamics from historical sales sequences and (ii) a Transformer encoder branch that models supplementary auxiliary features and statistically derived anomaly signals computed from that same sales series, including STL residual–based indicators, change-point flags, and calendar variables. The fused representation is jointly optimized to predict both the demand two days ahead of the input window and the probability of a statistically defined anomaly event on that day. Experiments are conducted on the public Kaggle Store Item Demand Forecasting dataset (10 stores, 50 items, daily sales from 2013 to 2017), using the most recent 90 days in each series for validation and synthetic anomaly labels derived from statistical residual analysis. Compared with a unimodal LSTM baseline, the proposed model reduces RMSE by approximately 7.23% and MAE by approximately 6.86%, and the bootstrap confidence intervals of both gaps exclude zero. Under an identical evaluation protocol the model also outperforms a naive forecast, a seasonal naive model and AutoARIMA, whereas a global gradient-boosted tree baseline attains a slightly lower point error (RMSE 0.5574 versus 0.5798). Because ground-truth incident labels are unavailable for this dataset, the classification task is supervised by weak labels derived from the residuals of the same series; the resulting discrimination is limited (ROC–AUC between 0.578 and 0.589 and PR–AUC between 0.126 and 0.141 against a positive rate of 0.100), and the anomaly head is therefore presented as a calibrated, cost-aware screening signal rather than as a validated anomaly detector. Under a 5:1 false-negative-to-false-positive cost ratio the proposed model attains the lowest expected cost among the configurations compared, and temperature scaling yields well-calibrated alert probabilities, so that a single multitask model can support both replenishment decisions and screening-level early warning.