Murphy, 1988) and, in transportation, the decision on how much to invest in snow removal equipment (Howe and Cochrane, 1976).
The two crop choice studies cited above (Tice and Clouser, 1982; Adams et al., 1995) are illustrative of the issues concerning rational expectations raised by Johnson and Holt (1997). Tice and Clouser (1982) examined the use of a seasonal climate forecast by a farmer to determine relative area planted to corn versus soybeans. The forecast was assumed to be perfectly accurate. Two allocations of crops are computed by using historical climatic averages and by using forecast information. The simple arithmetic difference between average net revenues per hectare using climatic averages and that dictated by the forecast was computed. Use of the forecast to allocate areas planted in corn and soybeans was shown to increase revenues by $3.65 per hectare per year beyond revenues using historical climate.
Adams et al. (1995) investigated the use of climate forecasts to determine allocations of areas planted to cotton, corn, sorghum and soybeans in the southeastern United States. The chief difference between their study and that of Tice and Clouser (1982) was that they computed forecast value in terms of total net social welfare (combined producer and consumer surplus) for the nation rather than revenues for the individual farmer. Using a general equilibrium economic model, they computed welfare using forecast-assisted crop allocations under an assumption that all southeastern farmers would plant accordingly. Furthermore, they explicitly considered the case in which forecast accuracy is imperfect. They found that the use of a perfect forecast increased social welfare by $145 to $265 million per year. The use of an imperfect (though still skillful) forecast increased welfare by $96 to $130 million per year.
Several research problems remain unsolved for Bayesian decision theory applications to climate forecasts. These applications do not address how forecast information available in an invariant, and possibly irrelevant, format is made relevant and incorporated into individual decision makers' information requirements, which differ considerably from one decision maker to the next. They do not adequately explore the possibility that decision makers' utility functions are nonlinear. Most applications do not estimate the distributional effects of the use of forecasts (i.e., winners versus losers). Finally, the lack of data and empirical techniques for clearly valuing forecasts precludes the testing of Bayesian models against the real world.
There remain some significant challenges in applying the general concept of the value of forecasts. One is in addressing the imperfections in