Regime‑aware causal machine learning markedly improves US macro forecasts: Double Machine Learning gives the most robust out‑of‑sample accuracy, and explainability tools show uncertainty, financial volatility, energy shocks and monetary policy dominate inflation during crises.

Main Finding

An integrated causal machine learning framework that combines structural-break and regime detection with causal ML estimators and XAI substantially improves macroeconomic forecasting in unstable environments. Double Machine Learning (DML) delivers the best and most robust out-of-sample forecasts across horizons, regimes, and robustness checks, and causal/XAI analysis shows that uncertainty measures, financial volatility, oil-price shocks, and monetary policy dynamics become dominant drivers—especially for inflation during crises.

Key Points

• Problem: Traditional econometric models (VAR, TVP-VAR) and off‑the‑shelf ML models degrade sharply when parameters shift, dynamics are nonlinear, or uncertainty is high.
• Framework: A unified pipeline merges structural-break detection (Bai–Perron), Markov-Switching regime identification, causal ML estimation (Double Machine Learning and Causal Forest), and explainability (SHAP).
• Benchmarks: Compared to VAR, TVP‑VAR, Random Forest, XGBoost, and LSTM, the causal ML approach (DML in particular) gives superior forecasting accuracy.
• Regime dependence: Macro relationships are highly regime-dependent; coefficients and predictor importance change across identified regimes (normal vs crisis).
• Dominant drivers: Uncertainty indicators, financial volatility, energy-market variables (oil shocks), and monetary-policy measures strongly affect forecast performance and become especially salient in crisis regimes.
• Explainability: SHAP analysis reveals that uncertainty and energy-market variables drive inflation forecasts during high-instability episodes, improving interpretability for users.
• Robustness: Performance gains hold across multiple forecasting horizons, regime specifications, and robustness checks reported in the study.
• Contribution: Bridges econometric forecasting theory, causal-inference methods, ML algorithms, and XAI into a single empirical architecture aimed at more adaptive, transparent forecasting.

Data & Methods

• Data: U.S. macroeconomic time series drawn from publicly available databases; includes inflation, monetary policy indicators, financial variables, energy-market variables, and multiple uncertainty measures; sample includes major crisis episodes (2008 GFC, COVID‑19, 2022 inflation shock).
• Structural/regime detection:
  • Bai–Perron tests for multiple structural breaks to identify break dates and sample partitions.
  • Markov‑Switching models to characterize latent regimes (e.g., low- vs high-volatility states).
• Causal estimation and forecasting:
  • Double Machine Learning (DML) to estimate causal/structural relationships while flexibly controlling for many covariates.
  • Causal Forests to allow heterogeneous treatment effects / conditional relationships across observations and regimes.
• Explainability: SHAP values used to attribute feature importance and observe how drivers change across regimes and time.
• Benchmarks & evaluation:
  • Benchmarked against VAR, TVP‑VAR, Random Forest, XGBoost, and LSTM.
  • Forecast performance evaluated out-of-sample across multiple horizons and regime-specific subsets (typical metrics: RMSE/MAE and formal forecast-comparison tests).
• Robustness checks: alternative break/regime specifications, variable sets, and training/validation splits to test stability of findings.

Implications for AI Economics

• For modelers and central banks:
  • Adopt regime-aware and causal ML methods (e.g., DML, Causal Forest) to improve resilience of forecasts during crises and structural change.
  • Integrate structural-break detection and regime identification as routine pre-processing steps rather than assuming parameter stability.
  • Use XAI (SHAP or similar) to make ML-based forecasts interpretable and actionable for policy — especially important for communicating drivers of inflation and risk to stakeholders.
• For policy and decision making:
  • Place greater weight on uncertainty indicators, financial volatility, and energy shocks in real-time monitoring and policy reaction functions during unstable periods.
  • Prefer forecasting systems that explicitly model heterogeneous and regime-dependent causal effects when designing contingency policies.
• For research and deployment:
  • Blend econometric identification strategies with flexible ML to obtain both credible causal estimates and strong predictive performance.
  • Ensure robust out‑of‑sample validation across regimes; remain vigilant to data‑snooping/overfitting risks when integrating many covariates.
  • Consider computational and implementation costs: causal ML + regime detection requires more computation and careful tuning than standard models.
• Limitations and open questions:
  • Sensitivity to break/regime detection choices and sample period; transferability to other countries or smaller samples needs testing.
  • Identification assumptions underlying causal ML (e.g., unconfoundedness given controls) should be carefully assessed in macro settings.
  • SHAP and other XAI methods help interpret importance but do not replace careful economic interpretation of mechanisms.
• Policy/regulatory considerations:
  • Transparent, explainable causal ML systems could bolster trust and adoption by policymakers and regulators, but require clear documentation of assumptions, training data, and regime‑detection rules.

Overall, the study provides evidence that combining structural-break/regime methods with causal machine learning and explainability yields more accurate, robust, and interpretable macro forecasts—an approach with clear practical value for AI-driven forecasting in macroeconomics.