Predicting Invoice Dilution in Supply Chain Finance with Leakage-Free Two-Stage Models: XGBoost, KAN (Kolmogorov–Arnold Networks), and Ensembles

Battle Count: 75

The paper is primarily focused on supply chain finance risk management rather than quantitative trading. However, the two-stage classification-regression framework, ensemble methods, and leakage-free time-series feature engineering are transferable techniques relevant to financial risk modeling. The macroeconomic feature integration and rolling-window evaluation methodology have some applicability to trading signal generation, but the core problem (invoice dilution prediction) is not directly related to trading strategies.

The two-stage architecture with multiple regressor families and ensemble combinations requires moderate-to-high implementation effort. Leakage-free feature engineering across multiple lookback horizons (180/360/720 days, all-time) for buyer and buyer-supplier pairs adds complexity. Rolling-window retraining across seven windows, hyperparameter optimization for multiple model families, and the integration of macroeconomic indicators increase engineering overhead. However, individual components (XGBoost, RandomForest, MLP) are well-supported by standard ML libraries. FasterKAN and the ScoreAI pipeline may require custom implementation.

The paper describes the methodology, dataset schema (9 core fields, 137 engineered features), and evaluation protocol (7 rolling windows, 70/15/15 split) in detail. However, the production dataset from The Interface Financial Group is proprietary and not publicly available. No code repository is provided. Hyperparameter optimization details are mentioned but not fully specified. The ScoreAI pipeline is referenced but not open-sourced.