HybridStack: Interpretable Macro-Financial Forecasting Framework
Published:
Macro-Financial Forecasting Hybrid Ensembles & Bayesian Optimization SHAP & LIME Interpretability Published in Elsevier Array (Q1)
Project Overview
HybridStack is an explainable machine learning forecasting engine engineered to predict macroeconomic indices and high-frequency inflation dynamics with extreme numerical fidelity and transparent attribution.
Architectural Innovations
- Multi-Tier Ensemble Architecture: Blends gradient-boosted decision trees (
XGBoost,LightGBM,CatBoost) feeding into regularized meta-regressors (Ridge, ElasticNet) to prevent overfitting across volatile economic regimes. - Bayesian Hyperparameter Search: Multi-dimensional Bayesian tuning evaluating over 10 high-frequency macro-financial time-series sources.
- Superior Predictive Accuracy: Recorded an empirical held-out test RMSE of 0.0006, outstripping conventional econometric models (ARIMA, VAR) and deep recurrent baselines.
- Policy-Grade Explainability: Leverages TreeSHAP, Kernel SHAP, LIME, and Partial Dependence Plots (PDP) to quantify exact marginal feature impacts, providing institutional transparency for central banks and economic policymakers.