Table 2 Summary of machine-learning models used for EOL demand forecasting.
Model | Type | Core assumption/learning principle | Key strengths | Limitations in EOL context |
|---|---|---|---|---|
Elasticnet | Linear regression with L1 + L2 regularization | Demand follows quasi-linear trend; correlated features penalized proportionally | Handles multicollinearity; interpretable coefficients | Cannot capture non-linear or discontinuous demand behaviour; sensitive to zero-inflated data |
Huber regressor | Robust linear model combining least-squares and absolute loss | Outliers are limited by Huber threshold; assumes smooth residual distribution | Resistant to noise and extreme values | Linear formulation underfits sparse, highly non-stationary data |
Random forest | Ensemble of decision trees using bootstrap aggregation | Non-parametric; learns non-linear interactions without feature scaling | High accuracy, low variance; interpretable feature importance; stable on sparse data | Large memory footprint; limited extrapolation beyond training range |
XGBoost | Gradient-boosted tree ensemble | Sequential residual correction; regularized boosting for overfitting control | Fast training; strong generalization; handles mixed feature types | Prone to overfitting on small datasets; sensitive to learning-rate tuning |
LightGBM | Leaf-wise gradient boosting with histogram optimization | Greedy leaf growth to minimize loss | Very fast on large datasets; efficient parallelization | May overfit or produce unstable forecasts under high sparsity |
Catboost | Ordered boosting with categorical feature encoding | Uses permutation-driven boosting and target statistics | Handles categorical variables efficiently; reduced overfitting bias | Sensitive to noise; can produce abrupt forecasts when historical trend shifts |
Proposed RF + decay blending | Random Forest combined with exponential-decay modulation | ML prediction dynamically adjusted by lifecycle decay kernel | Physically interpretable; robust long-horizon extrapolation; suitable for SKU-level deployment | Requires empirical tuning of blending coefficient; current version deterministic (to be extended with uncertainty-aware methods) |
