Using Dynamic Forecasting Genetic Programming (Dfgp) To Forecast United States Gross Domestic Product (Us Gdp) With Military Expenditure As An Explanatory Variable

Classic time-series forecasting models can be divided into exponential smoothing, regression, ARIMA, threshold, and GARCH models. Functional form is investigator-specified, and all methods assume that the data generation process across all segments of the examined time-series is constant. In contrast, the aim of heuristic methods is to automate the discovery of functional form and permit different segments of a time-series to stem from different underlying data generation processes. These methods are categorized into those based on neural networks (NN) and those based on evolutionary computation, the latter further divided into genetic algorithms (GA), evolutionary programming (EP), and genetic programming (GP). However, the duration of the time-series itself is still investigator determined. This paper uses a dynamic forecasting version of GP (DFGP), where even the length of the time-series is automatically discovered. The method is applied to an examination of US GDP that includes military expenditure among its determinants and is compared to a regression-based forecast. We find that DFGP and a regression-based forecast yield comparable results but with the significant proviso that DFGP does not make any prior assumption about functional form or the time-span from which forecasts are produced.