A general framework for analyzing multi-dimensional panel data of survey forecasts

In this dissertation, I define a model and develop a methodology for analyzing forecasts and forecast errors using three-dimensional panel data. Unlike traditional econometric models which attempt to fit the data to general functional forms, my model is based on equations which describe the process by which forecasts and targets are generated. These descriptive equations show precisely what functional form rationality tests should take and they imply direct measures for aggregate shocks and economic uncertainty, series which were heretofore unobservable.

The model allows for the complex correlation among forecasts made by different individuals, for different targets, and at different horizons to be expressed as a function of a few fundamental parameters. Applying the model to three-dimensional data, I achieve more comprehensive rationality tests by expressing the error covariance matrix as a function of these fundamental parameters and estimating these parameters directly from the data prior to testing. More importantly, my model directly gives measures of aggregate shocks impacting specific economic series prior to the series being realized.

My model also gives measures of the volatility of the aggregate shocks falling over time, measures of the anticipated changes in economic series, measures of the volatility of the anticipated changes, and measures of the uncertainty of each forecaster independent of the aggregate shocks at each point of time. In addition, I formulate a methodology for dealing with randomly missing observations in panel data. I apply my model to data from the Blue Chip and ASA-NBER forecast surveys and to data from the ASA-NBER probability forecast survey.

Applying the model to probability forecasts, I find that traditional proxy measures of forecast uncertainty taken from point forecasts are inadequate. My model shows the precise formulation for uncertainty as well as the traditional proxy in terms of fundamental underlying parameters. Empirical evidence also indicates that traditional ARCH modelling of uncertainty is inadequate.