Overview
Tom Stark was a research economist at the Federal Reserve Bank of Philadelphia. His work focused on Bayesian VAR forecasting methodology for U.S. macroeconomic variables, with particular emphasis on incorporating cointegrating restrictions within a Bayesian shrinkage framework. His 1998 working paper remains one of the most detailed real-time evaluations of a Bayesian VEC specification, covering a 22-year rolling forecasting experiment (mid-1975–1997Q4).
Key Contributions / Features
- BVEC model (Working Paper 98-12, FRBP, 1998): Specified and evaluated a 7-variable Bayesian Vector Error Correction model for short-term forecasting of U.S. output growth, inflation, and unemployment. Demonstrated that explicit differencing and the error correction term jointly improve forecast performance relative to naive BVAR alternatives.
- Modified Litterman prior: Departed from standard Litterman in two ways — (1) imposed unit roots by explicit differencing rather than assigning a random-walk prior with finite variance; (2) used lag decay exponent γ=0.50 rather than the standard γ=1.0, allowing more distant lags to contribute information to the posterior.
- Diffuse prior on EC coefficient: Following LeSage (1990) and Joutz-Maddala-Trost (1995), assigned a non-informative prior to the error correction loading while retaining informative Minnesota priors on the short-run dynamics. Estimated via Theil's mixed estimator.
- Fisher relation evaluation: Demonstrated that imposing real rate stationarity (the Fisher relation) as a second EC term worsens 2-year inflation forecast accuracy (RMSE rises from 1.52% to 2.36%), providing concrete evidence against enforcing the Fisher effect as a long-run cointegrating constraint in forecasting systems.
- Bayesian priors essential: Table 6 comparison with OLS shows that removing priors raises 1-year inflation RMSE from 0.93% to 1.35%, GDP growth from 1.82% to 2.60%, and unemployment from 0.52% to 0.72%.
Related