Scholarlink

<p>This study forecasts global economic policy uncertainty using deep statistical learning models applied to monthly data from 1997 to 2024. The forecasting target is the stationary monthly change in the Global Economic Policy Uncertainty index, while the index level itself is treated as highly persistent. Compact one-dimensional CNN and LSTM architectures are estimated using fixed input windows and are benchmarked against ARIMA and GARCH models under a strict date-based train-test split and rolling-origin cross-validation design. Hyperparameters are optimized using Adam-based methods and Differential Evolution, and model performance is evaluated on common holdout samples using RMSE, MAE, MAPE, and R-squared. The best-performing CNN model, optimized with Adam, achieves out-of-sample RMSE improvements of approximately 16% to 18% relative to ARIMA and GARCH and also outperforms the LSTM model in terms of MAE, MAPE, and R-squared. In addition, a dedicated volatility head learns conditional variance from recent temporal sequences and generates regime-dependent 80% and 95% prediction intervals that widen during periods of stress and narrow during calmer episodes. The forecasts indicate gradual mean reversion in global policy uncertainty, interrupted by occasional spikes that coincide with major historical stress events. The study contributes a unified and computationally efficient forecasting architecture that combines accurate point prediction with calibrated uncertainty bands for macro-financial risk assessment and is fully reproducible through an open code repository and well-documented implementation procedures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advanced Deep Statistical Learning Approach for Forecasting Global Economic Policy Uncertainty and Volatility

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Khder Alakkari,
  • Mostafa Abotaleb,
  • El-Sayed M. El-kenawy,
  • Pradeep Mishra

摘要

This study forecasts global economic policy uncertainty using deep statistical learning models applied to monthly data from 1997 to 2024. The forecasting target is the stationary monthly change in the Global Economic Policy Uncertainty index, while the index level itself is treated as highly persistent. Compact one-dimensional CNN and LSTM architectures are estimated using fixed input windows and are benchmarked against ARIMA and GARCH models under a strict date-based train-test split and rolling-origin cross-validation design. Hyperparameters are optimized using Adam-based methods and Differential Evolution, and model performance is evaluated on common holdout samples using RMSE, MAE, MAPE, and R-squared. The best-performing CNN model, optimized with Adam, achieves out-of-sample RMSE improvements of approximately 16% to 18% relative to ARIMA and GARCH and also outperforms the LSTM model in terms of MAE, MAPE, and R-squared. In addition, a dedicated volatility head learns conditional variance from recent temporal sequences and generates regime-dependent 80% and 95% prediction intervals that widen during periods of stress and narrow during calmer episodes. The forecasts indicate gradual mean reversion in global policy uncertainty, interrupted by occasional spikes that coincide with major historical stress events. The study contributes a unified and computationally efficient forecasting architecture that combines accurate point prediction with calibrated uncertainty bands for macro-financial risk assessment and is fully reproducible through an open code repository and well-documented implementation procedures.