A Crisis-Conditioned Reinforcement Learning Framework Demonstrates That Uncertainty-Aware Reward Design Is Necessary, Not Optional, for Global Macroeconomic Policy Optimization
Year: 2026
Type: preprint
Abstract: This paper introduces MAKOTO (Multi-domain Adaptive Knowledgefor Open-economy Trend Optimization), a crisis-conditioned deepreinforcement learning framework for global macroeconomic policyoptimization. MAKOTO embeds a continuous IsolationForestuncertainty score calibrated on a pre-crisis reference period(2004-2007) as a multiplicative amplifier within a ProximalPolicy Optimization (PPO) reward function. Applied to a30-country, 24-year World Bank panel (720 observations,2000-2023), MAKOTO achieves a statistically significant 7.1%total reward advantage over an unconditioned baseline (bootstrap95% CI: [+1.947, +4.154]; p < 0.05), robust across threeindependent random seeds, and translating to an estimated $18.8billion in annual welfare gains. An ablation study establishesthat the uncertainty signal is structurally integral to thelearned policy: removing it at evaluation time causesperformance to fall below the unconditioned baseline, confirmingthat crisis-conditioned reward architecture is a necessarydesign principle, not an additive enhancement.
Author: Nubaid Khan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162543293
Cites:
Cited by:
Related to: