Multi-Layer Ensemble Decision-Tree

This model is a regime/structural-change detection layer intended to reduce overconfidence under distribution shift and improve discretionary calibration. The model is built as a multi-layer ensemble with a decision-tree / rule-learning backbone (RuleFit-style) augmented by heterogeneous learners (e.g., gradient-boosted models and probabilistic decision-process components). Its differentiator is the training and evaluation setup: it is trained in a zero-shot forecasting spirit across heterogeneous domains-market series plus non-market series (weather, energy, ECG-like signals-to enforce invariances and penalize “market-only shortcuts.” The pipeline emphasizes feature hygiene and interpretability: signal-to-noise screening, residualisation/whitening/normalization, redundancy controls, and partial-dependence-driven checks. The objective is a detector that fires on structural change patterns (breaks, phase shifts, observability loss) and outputs actionable calibration signals (risk throttles, model confidence caps, or regime labels) suitable for downstream discretionary or systematic decision rules.