Performance of artificial neural networks in forecasting Costa Rican inflation

Resumen

The underlying processes that generate economic series such as inflation, unemployment or output gaps are potentially quite complex. Undoubtedly that makes it very difficult to forecast them and has traditionally bent attention to relatively simple linear approaches when trying to model them. Trying to capture nonlinear relationships among inflation and its determinants, this paper applies Artificial Neural Networks (ANN) to forecast Costa Rican inflation.  A innovative technique that systematically discriminates among different networks in order to overcome the problem of “over-fitting” a ANN was applied.  Forecasts are compared with those obtained from “thick” models and traditional linear techniques. The potentially complex nonlinear relationships between inflation and its short run determinants in an expectations-augmented Phillips Curve scheme are captured with a systematically chosen ANN.  Forecasts at different horizons are computed in a rolling exercise in order to test the hypothesis of a better performance of the nonlinear parameterization.  Evidence shows that linear techniques do not outperform ANN and, in the case of a Phillips Curve, networks forecasts statistically improve upon linear approaches especially for short run forecast horizons.  In most cases, “thick” modeled ANN’s forecasts showed a weak performance compared with systematically chosen ANNs.