سيما اميري زاده

Accurate forecasting of backorders is crucial for e-commerce operations to minimize inventory, reduce lost sales, an‎d optimize logistics costs. This thesis develops an‎d rigorously eva‎luates a hybrid machine learning architecture that leverages the complementary strengths of LSTM, Transformer, XGBoost, an‎d logistic regression models to predict backorders under highly nonlinear an‎d autocorrelated deman‎d an‎d inventory conditions. An exploratory analysis revealed poor distributions in key volume characteristics an‎d weak linear correlations between inventory an‎d deman‎d, indicating the inadequacy of traditional statistical methods. By integrating LSTM temporal pattern learning, Transformer sequence attention, XGBoost nonlinear feature interaction capture, an‎d logistic regression interpretability, the proposed system achieved an overall accuracy of 85.1%, an F-1 score of 84.3%, a precision of 88.6%, an‎d a recall of 80.4%, which surpasses the performance of each individual model an‎d confirms that a multi-model approach significantly improves the reliability of the forecast. These results demonstrate the practical value of this framework for predictive inventory management an‎d emphasize its generalizability to real-time applications.

سفارش معوق , LSTM , XGBOOST , Transformer , جنگل تصادفي , يادگيري گروهي

Backorder , LSTM , Transformer , XGBOOST , Random Forest , Ensemble Learning

Sima Amirizadeh

Siamak Nouri