This Paper uses multi-step forecasting models at horizons of 4 and 8 quarters to forecast and explain the growth of real per capita US GDP. In the modeling strategy, a priori sign restrictions play an important role. They are imposed not on impulse response functions but directly on the reduced form single or multi-step equations, unlike in recent work by Uhlig and Canova. This is possible because in this context, the reduced form inherits important structural sign properties; basically, that autonomous expenditure has positive effects on near future GDP. We consider an economically large class of variables, including effects from interest rates, the credit channel and asset prices, the real exchange rate, yield spreads, inflation and interest rate volatility, oil prices (including asymmetries), structural breaks in fiscal and monetary policy, the recent behaviour of consumption, investment and profitability, and the evolutionary effect of globalization on the balance of payments constraint. We follow a general to specific methodology, including the help of PCGETS (Hendry and Krolzig, 2001) to reduce general models to more parsimonious ones. Relative to conventional VARs, our models imply longer lag structures than ever considered in VARs, as well as non-linearities, and so could never have been found with conventional VAR restrictions. Our results thus contradict the suggestion of Sims (1980) that VARs can resolve the problem of ‘incredible restrictions’ embodied in large macro econometric models. Our exercise of learning from the data through general to specific modeling is likely, in many cases, also to contradict the lag structures of such models. We present a range of models with remarkable recursive forecasting performance since 1982 and show that similar models could have been selected with 1982 data by applying similar methods then. Out of sample forecasts with such models since March 2001, when we forecast that 2001 would be a recession year, have also been successful.
