An Empirical Evaluation of Some Long-Horizon Macroeconomic Forecasts

We use long-run annual cross-country data to evaluate pseudo out-of-sample forecasts of five variables for horizons up to 50 years. The variables we forecast are real per capita GDP growth, CPI inflation, labor productivity growth, and long- and short-term nominal interest rates. Our models for forecasting include simple time series models and frequency domain methods recently developed in Müller and Watson (2016). We focus on coverage of 68% forecast intervals (that is, coverage of 68% confidence intervals for forecasts). For GDP growth, CPI inflation and labor productivity growth, median coverage across countries is roughly 68% for several models, but with considerable dispersion around that median. For these three series, a reasonable model choice is a frequency domain model that does not require the user to take a stand on the order of integration of the data. For interest rates, forecast intervals for most models and samples include markedly fewer than 68% of realized values. For interest rates, a reasonable model choice is a driftless random walk. For real per capita GDP and labor productivity growth, we find that forecasts and forecast intervals from the best-performing models are very similar to the Social Security Administration’s (SSA’s) long-run projections. In contrast, for CPI inflation and long-term interest rates, we find that forecasts from the best-performing models have wider forecast intervals than intervals implied by the SSA’s projections for their low- and high-cost scenarios