Abstract:
Inflation forecasting is crucial for efficient monetary policy and decision-making in an economy.
This paper examines the feasibility of including deep neural networks in the macroeconomic forecasting
toolbox for the South African economy. This study focuses on South Africa’s annual headline
inflation rate and applies two different deep neural network architectures for forecasting. The
deep neural network’s performance is compared to the autoregressive integrated moving average
(ARIMA) benchmark, where root mean squared error (RMSE) is used as a performance measure.
The results show that the multiple layer perceptron (MLP) outperformed the benchmark and
its peer, the convolutional recurrent neural network model. Admittedly, the convolutional long-short
term memory network (CNN-LSTM) is sensitive to architectural design, especially when the
amount of training data is in short supply. In conclusion, the study finds that the ARIMA model
predicts inflation inconsistently in the presence of endogenous and exogenous structural breaks in
the time series and consequently gives non-unique forecasts. The MLP becomes a viable addition
to the macroeconomic forecasting toolbox in such a case.
