A1006
Title: AI meets fiscal policy: Fiscal actions across 140+ countries
Authors: Adrian Peralta Alva - IMF (United States) [presenting]
Davide Furceri - IMF (United States)
Nikhil Patel - IMF (United States)
Shuvam Das - IMF (United States)
Abstract: A novel quarterly database of discretionary fiscal measures is presented for 143 economies over 1950-2024. Policy passages are drawn from Economist Intelligence Unit country reports and processed in two steps. First, a rule-based parser identifies candidate text segments covering fiscal developments. Second, a large language model (GPT 4.1) classifies each episode by (i) net fiscal stance (expansion, contraction, or neutral) (ii) quantitative scale, and (iii) policy motivation. In the subsample that overlaps a prior study, agreement rates are 92 percent for the sign of policy actions and 90 percent for motivations behind the policies. The database extends narrative coverage by an order of magnitude in both the cross-section and the time dimension. Descriptive statistics document the distribution of tax and expenditure shocks across regions, income groups, and periods. As an illustration, country-specific Bayesian VARs yield four-quarter output multipliers for exogenous fiscal consolidations of 0.7 in the United States and 1.5 in Botswana, consistent with heterogeneous transmission predicted by theory.
