Forecasting Costa Rican inflation with machine learning methods

We present a first assessment of the predictive ability of machine learning methods for inflation forecasting in Costa Rica. We compute forecasts using two variants of k-nearest neighbors, random forests, extreme gradient boosting and a long short-term memory (LSTM) network. We evaluate their properties according to criteria from the optimal forecast literature, and we compare their performance with that of an average of univariate inflation forecasts currently used by the Central Bank of Costa Rica. We find that the best-performing forecasts are those of LSTM, univariate KNN and, to a lesser extent, random forests. Furthermore, a combination performs better than the individual forecasts included in it and the average of the univariate forecasts. This combination not biased; its forecast errors show appropriate properties, and it improves the forecast accuracy at all horizons, both for the level of inflation and for the direction of its changes.