Comparing Artificial Neural Network and Cohort-Component Models for Population Forecasts

Abstract

Abstract:Artificial neural network (ANN) models are rarely used to forecast population in spite of their growing prominence in other fields. We compare the forecasts generated by ANN long short-term memory models (LSTM) with population projections from the traditional cohort-component method (CCM) for counties in Alabama, USA. The evaluation includes projections for all 67 counties, which are diverse in population and socioeconomic characteristics. When comparing projected values with total population counts from the 2010 decennial census, the CCM used by the Center for Business and Economic Research at the University of Alabama in 2001 produced comparable or better results than a basic multi-county ANN LSTM model. Results from ANN models improve when we use single-county models or proxy for a forecaster’s experience and personal judgment with potential economic forecasts. The results indicate the significance of forecaster’s experience/judgment for CCM and the difficulty, but not impossibility, of substituting these insights with available data.