ERIC Number: ED107709
Record Type: RIE
Publication Date: 1973-Jul
An Experiment in Probabilistic Forecasting.
Brown, Thomas A.
Students were asked to make forecasts of fourteen quantities where true values would not become known for five or six months. The quantities were selected to be typical of the subjects which would be of interest to a decisionmaker in business or government, and included GNP, consumer prices, draft calls, deaths in South Vietnam, and election results on both the state and national level. Rather than giving a one-point estimate for each quantity, each respondent was asked to give a probability distribution which reflected the likely behavior of the quantity in question. It was found that almost all respondents were able to give meaningful distributions and there was a tendency for the true answer to occur disproportionately often in the tails of the distribution given. The effect of this type of error can be counteracted to a certain extent by combining the individual responses into a "consensus" distribution, which will have a greater spread than most of the individuals' distributions. Of the four ways in which this combining of responses was carried out, the most effective was to average the individual's probability density functions. All four consensus methods produced better forecasts than did the average individual. No clear-cut association was found between forescasting skill and age, sex, academic major, or score on a simple current events quiz. (Author/BJG)
Publication Type: Reports - Research
Education Level: N/A
Sponsor: Advanced Research Projects Agency (DOD), Washington, DC.
Authoring Institution: Rand Corp., Santa Monica, CA.