A Hybrid Deep Learning and Machine Learning Framework for Quantifying the Macroeconomic Impact of Cyberattacks: Threat Intelligence, Vulnerability Assessment, and Predictive Resilience Modeling in National Economic Systems

Abstract

Cyberattacks on energy, finance, health and logistics systems now produce losses large enough to register at the level of national economies, yet the quantitative tools that link threat intelligence to macroeconomic exposure remain fragmented. This paper develops a hybrid framework that couples advanced econometrics with deep and machine learning across three modules: threat intelligence, vulnerability assessment and predictive resilience. Using a compiled dataset of published series for 2015–2025 (NVD/CVE disclosures, FBI IC3 complaints and losses, IBM/Ponemon breach costs), ITU Global Cybersecurity Index scores for 25 economies and 24 documented macro-relevant incidents, we estimate HAC and wild-bootstrap elasticities, ARDL bounds and DOLS long-run relations, Bayesian conjugate regressions, sup-F break and Granger tests, quantile, Tobit and fractional-logit models, and a Poisson–lognormal loss distribution with extreme-value diagnostics. A from-scratch LSTM deep ensemble, deep MLP, random forest, gradient boosting and an ARX–LSTM hybrid are evaluated by rolling-origin forecasting with Diebold–Mariano tests and exact Shapley attribution. Reported losses grew 35.6% a year against 22.4% for vulnerability disclosures; disclosures Granger-lead losses. Gradient boosting gave the lowest out-of-sample error (log RMSE 0.113 against 0.362 for a random walk); the LSTM ensemble reached 0.191. The hybrid projection puts US-reported losses at USD 52.0 billion by 2028 (80% interval 41.0–65.9). Incident severity is lognormal with a heavy right tail, and parameter uncertainty roughly doubles the 99% Value-at-Risk. Projected expected losses range from 0.15% of GDP in the United States to 0.26% in Nigeria, with Pakistan and Ukraine close behind.