Testing Exchange Rate Models in a Small Open Economy: an SVR Approach

We empirically test the validity of four popular monetary exchange rate models under five alternative inflation expectation approximations using the NOK/USD exchange rate. The selection of Norway seems appropriate as it is a small open economy that does not participate in most economic or political organizations and uses the Government Pension Fund as a tool to dampen external shocks to the domestic economy. The main innovation of the paper is that in addition to a standard VECM model used in the literature, we employ a two-step procedure for the first time in this setting; first, we train a Support Vector Regression (SVR) model and then we extract the coefficients through a Dynamic Evolving Neural Fuzzy Inference System (DENFIS). The best overall model in terms of fitting the phenomenon is an SVR one with autoregressive inflation expectations that exclude energy prices, exhibiting four times lower forecasting error than the best VECM model. The estimated coefficients of the VECM are not statistically significant, while the ones from the SVR-DENFIS model show that the sign of the coefficient on the interest rate differential corroborates only with the model proposed by Bilson (1978), while we detect a significant inflation rate differential. We conclude that fundamentals possess adequate forecasting ability when used in exchange rate forecasting but none of the tested monetary exchange rate models can explicitly describe the evolution path of the exchange rate. Nevertheless, the proposed machine learning methodology moves one step further than the econometric approach in tackling the exchange rate disconnect puzzle.