More Is Better: Consensual Forecasting and State Revenue Forecast Error

Abstract

The subject of this research is the question of whether revenue forecasts, which require consensus within the institutional framework of federal system governments, are more accurate than in states where forecast decision-making rests solely in the executive branch or in the legislative branch. Results are reported from a 1999 survey of various revenue forecasters in the 50 states. The methodology includes considerations of the following as independent variables: split-government legislatures; frequency of state's forecasts; whether a separate council of economic advisors was included; budgetary balance requirements; the availability of outside expert advice from universities; and the extent of a period of economic stability in the state. The results indicate some effect, which although not substantial, in terms of state budgets still constitute significant dollar amounts.

Notes

^aThe Shapiro–Wilk test35 performs a check for normalcy of a distribution. The Shapiro–Wilk statistic is .688 and thus the distribution can be rejected as a normal distribution at a 99% confidence level. After transformation, a new Shapiro–Wilk statistic is generated showing a value of .99674 (1 is considered the mean value for a normal distribution) and cannot be rejected as normal at the 90% confidence level. To make the following pages less verbose the “log of the forecast error” or the “log of the error” will be used as a shorthand for the transformed dependent variable that is actually the ln(forecast error + .5).

^bIt begins by estimating the equation y _i  = x _i b + z _i t + u where z takes on the values of . A standard F-test is then performed to determine if t = 0. The results show an F-statistic of 0.82 with 3318 degrees of freedom.

^cThis test looks for heteroscedasticity by modeling the variance as a function of fitted values. The results of the Cook–Weisberg test indicate that a nonconstant variance or heteroscedasticity can be rejected based on a chi-square statistic of 0.84 with one degree of freedom.