Forecasting the U.S. macroeconomy

Most economic forecasts are reported as a single number, yet the decisions that depend on them turn on uncertainty and timing, not on a point estimate. Our latest working paper builds a transparent, fully reproducible forecasting engine that combines four standard models, a Bayesian vector autoregression with stochastic volatility, a factor model, a regime-switching model, and a machine-learning component, into one probability distribution for six U.S. series: CPI and core PCE inflation, industrial production, payrolls, jobless claims, and the unemployment rate. We test it the way a user would, asking whether its stated uncertainty is honest and whether it leans the right way before a turn. Evaluated in real time on a decade of unrevised data, the combined forecast is well calibrated, cuts error against a random walk by up to half, and warns in advance when its own errors are likely to be large. It is also candid about its limits: for a near-random-walk series like the unemployment rate, we show plainly that no combination adds much. The result is a calibrated, auditable second opinion for anyone who needs a forecast they can defend. Read the paper here.