Error Reduction strategies for the 1998-2005 USAGE Forecast

This paper examines methods aimed at improving baseline economic forecasts using a dynamic CGE model. Forecasting can be used to test the validity of such models, as well as to highlight possible improvements, by investigating the discrepancies between the forecast and actual outcomes. The model employed here is USAGE - a recursive dynamic, 500-industry CGE model of the U.S. USAGE generates baseline forecasts by incorporating expert forecasts for certain macro variables and extrapolating historical trends in technology, consumer preferences, positions of foreign demand curves for U.S. products, and numerous other naturally exogenous variables. In instances where important trends either dissipate or reverse, large forecast errors can arise. This paper seeks to provide explanations and guidance as to whether these various trends from the period 1992 to 1998 would continue for the 1998 to 2005 USAGE forecast. The twenty largest errors on a relative and/or absolute basis are examined. It is found that for some commodities, had all publicly available information by 1998 been appropriately utilised, certain important trends should not have been expected to continue. Hence, a better forecast could have been generated had the projection of certain trends been nullified. More generally, the findings suggest that there is a case to be argued against projecting forward large values relating to import-domestic preference twist factors in particular. It is also shown that for commodities in the trade-exposed textile, clothing and footwear industries moderately better results could have been produced by implementing import price forecasts in a way that is more in line with historical trade policy. This was achieved by projecting forward real basic import prices. However, the key drivers behind these errors were usually the significant underestimation of the impact of import-domestic preference twist factors, as well as the overestimation of factor input cost savings. In relation to forecasts for commodities in the oil and mining sectors as well as industries that service these cyclical industries, it is concluded that these typically could not have been improved in the absence of strong convictions (by 1998) about an impending mining "super-cycle" or extended boom. For the construction-related commodities demand was fuelled by virtually unprecedented low borrowing costs. In these instances, it is difficult to conclusively argue that the modeller could have produced a better forecast. Moreover, while large improvements in forecast accuracy can be obtained for some industries and sectors, the overall economy-wide forecast error does not fall greatly due to the sheer volume of commodities. While it is disappointing that the error is not very reducible, it is also reassuring because it implies that the default implementation of the model is quite powerful. In all about 4% of all commodities were specifically examined to assess the potential for error reduction. After due consideration about 7.5% of commodities were in some way directly re-projected. To generate a large reduction in the forecast error would require an extensive amount of work and probably call for the input of numerous industry specialists.