Inflation Rate Forecasting: Extreme Learning Machine as a Model Combination Method

Inflation rate forecasting is one most discussed topics on time series analysis due to its importance on macroeconomic policy. The majority of these papers findings point out that forecasting combination methods usually outperform individual models. In this sense, we evaluate a novel method to combine forecasts based on Extreme Learning Machine Method (Huang et al., 2004), which is becoming very popular but, to the best of our knowledge, has not been used to this purpose. We test Inflation Rate forecasting for four Latin American countries, for one, two, three, ten, eleven and twelve steps ahead. The models to be combined are automatically estimated by R forecast package, as SARIMA, Exponential Smoothing, ARFIMA, Spline Regression, and Artificial Neural Networks. Another goal of our paper is to test our model against classical combination methods such Granger Bates, Linear Regression and Average Mean of models as benchmarks, but also test it against basic forms of new models in the literature, like Diebold and Shin (2018), Garcia et al. (2017) and Wang et al. (2018b). Therefore, our paper also contributes to the discussion of forecast combination by comparing versions of some methods that have not been tested against each other. Our results indicate that none of these methods have an indisputable superiority against the others, however the Extreme Machine Learning Method proved to be the most efficient of all, with the smaller Mean Absolute Error and Mean Squared Error for its predictions.