Evaluating forecast accuracy: a comparative error-based analysis of ann and markov-switching models in inflation prediction

Abstract

In this study, we develop a hybrid modelling approach that integrates artificial neural networks with a Markov-switching autoregressive model to enhance the accuracy of inflation predictions, utilising monthly data from South Africa from 2009 to 2023. Ten statistical loss functions, including mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE), mean absolute scaled error (MASE), and Theil’s U, are used to measure how well the forecast works. The MS(2)-AR(1)-ANN(5,3) performed the best overall out of all the competing specifications. The model has a mean squared error of 0.9897, a mean absolute error of 0.7573, a root mean squared error of 0.9948, and a mean absolute scaled error of 0.115. The model further yields Theil’s U statistic of 3.568, where, out of the 100 loss functions used, it ranks first in eight, second in the mean prediction error, and third in the symmetric mean absolute percentage error, with an SMAPE score of 18.234. This score is slightly higher than some of the base learners, but scale-free measures like MASE are better at providing trustworthy advice when conditions change quickly. The findings show that percentage-based measures like MAPE have their limits and that MASE is a better predictor during times of structural change. In general, the results show that the hybrid MS-AR-ANN architecture produces inflation projections far more accurate than those from the separate base models. The suggested methodology offers valuable insights for policymakers and central banks seeking early-warning indicators and robust inflation-monitoring systems amid regime upheavals and economic turmoil.