This paper describes how the Congressional Budget Office uses a Bayesian vector autoregression (BVAR) method to generate alternative economic projections to the agency’s baseline. The BVAR includes a wide range of key economic variables that are needed to approximate budget outcomes. Its estimation methods avoid overfitting, a situation in which a model fits historical data well while having a poor ability to project future values.
Given targets of future values of some variables such as inflation, the BVAR generates economic projections consistent with the targets and historical dynamics of the variables in the model. The BVAR is a flexible framework that can incorporate new variables and impose conditions for alternative economic projections. It also has forecasting performance comparable to that of CBO’s baseline forecasting method.
