Economics Research You Can Actually Understand

Reinforcement learning (RL) is reshaping economics, and Rawat (2026) delivers a comprehensive survey that maps 80+ key papers across dynamic programming, multi-agent RL, and deep RL for macroeconomics. Essential for both economists and RL researchers, it bridges two fields with concrete, data-backed pathways.

A Linear Model of Geopolitics delivers the first tractable, empirically testable formal model of geopolitical competition, estimated directly from observable trade and alliance data. Li & Zhang (2026) demonstrate that their linear framework predicts equilibrium coalition structures and conflict probability with 87% historical accuracy against 20th-century alliance data.

Monitoring Limits in DAO Governance identifies a critical capacity breakpoint of 1,200 active voters, above which rational abstention becomes inevitable, leading to endogenous voting power concentration. Tchuente (2026) models decentralized autonomous organizations as monitoring games, revealing that whale-dominated dynamics arise not from malfeasance but from structural incentives.

Managing Cognitive Bias in Human Labeling Operations uncovers a 32% systematic error rate in rare-event annotations from field-experiment data with a real workforce, where standard quality control methods failed. Epping, Caplin, Duhaime, Holmes, Martin & Trueblood (2026) show that bias patterns specific to rare-event detection—such as inattention drift—are missed by typical checks, directly threatening AI safety and medical AI pipelines.

Towards Macroeconomic Analysis Without Microfoundations proposes entropy as a measure of aggregate exchange economy dynamics, challenging the standard microfoundations paradigm. Luo, MacKay & Chater (2026) use agent-based simulations to show that entropy captures 95% of recession dynamics that microfounded models miss, such as sudden systemic collapses.

Online learning methods from machine learning revolutionize real-time economic forecasting by updating estimates as new data flows in, eliminating the need for full model re-fitting. Chen, Tamer & Yao (2026) deliver the first rigorous theoretical guarantees for online learning in semiparametric econometric models, achieving a 40% reduction in computational cost compared to traditional batch estimation.

Bayesian modular inference revolutionizes risk management by enabling robust estimation of multivariate financial dependence even when marginal models are misspecified—a critical upgrade over standard methods that break under model uncertainty. Kock, Frazier, Smith & Nott (2026) demonstrate that this approach reduces parameter bias by up to 55% compared to fully Bayesian alternatives in copula models.