 |
This paper demonstrates that machine learning techniques can be used to efficiently forecast macroeconomic time series. We show that artificial neural networks outperform a linear autoregressive (AR) and a random walk (RW) models in forecasting the monthly US CPI inflation. One-step-ahead RMSE of a simple neural network (NN) and of a long short-term memory (LSTM) recurrent neural network is approximately half of the corresponding measure for the AR or RW models. For short horizons, up to 3 steps ahead, both NN and LSTM give more accurate forecasts than the AR and RW. At longer horizons, 6 through 12 steps ahead, the performance of the NN becomes on a par with the linear AR and RW models. However, the LSTM continues to produce more accurate predictions with the errors being approximately 40% smaller than of the RW or AR. Additionally we conduct a sensitivity analysis with respect to hyper-parameters and provide a qualitative interpretation of what the networks learn by applying a novel layer-wise relevance propagation technique.
