Why are forecasts of inflation from VAR models so much worse then their forecasts of real variables? This paper documents that relatively poor performance, and finds that the price equation of a VAR model fitted to U.S. postwar data is poorly specified. Statistical work by other authors has found that coefficients in such price equations may not be constant. Based on specific monetary actions, two changes in monetary policy regimes are proposed. Accounting for those two shifts yields significantly more accurate forecasts and lessens the evidence of misspecification.