Abstract:
Forecasting aggregate retail sales may improve portfolio investors‟ ability to predict movements
in the stock prices of the retailing chains. Therefore, this paper uses 26 (23 single and 3
combination) forecasting models to forecast South Africa‟s aggregate seasonal retail sales. We
use data from 1970:01 – 2012:05, with 1987:01-2012:05 as the out-of-sample period. Unlike, the
previous literature on retail sales forecasting, we not only look at a wider array of linear and
nonlinear models, but also generate multi-steps-ahead forecasts using a real-time recursive
estimation scheme over the out-of-sample period, to mimic better the practical scenario faced by
agents making retailing decisions. In addition, we deviate from the uniform symmetric quadratic
loss function typically used in forecast evaluation exercises, by considering loss functions that
overweight forecast error in booms and recessions. Focusing on the single models alone, results
show that their performances differ greatly across forecast horizons and for different weighting
schemes, with no unique model performing the best across various scenarios. However, the
combination forecasts models, especially the discounted mean-square forecast error method
which weighs current information more than past, produced not only better forecasts, but were
also largely unaffected by business cycles and time horizons. This result, along with the fact that
individual nonlinear models performed better than linear models, led us to conclude that
theoretical research on retail sales should look at developing dynamic stochastic general
equilibrium models which not only incorporates learning behaviour, but also allows the
behavioural parameters of the model to be state-dependent, to account for regime-switching
behaviour across alternative states of the economy.
