This study analizes the forecast ability of the sign variation of the daily exchange in neural networks. The forecast of price's direction is important to effective trade strategies (Leung, Daouk & Chen, 2000). We used backpropagation neural network models: the Multilayer Perceptron, the Jordan-Elman Recurrent network and the Ward network, all with standard and recursive method. To test the forecast ability we use an out-of-sample set of 398 observations and over 337 out-of-sample sub-set of 60 observations. The relative performance of the models was measured by the number of hits of the exchange rate variation sign, using for that the Pesaran & Timmermann s (1992) directional accuracy test. Then, the results of the best neural network model was compared with the ones of AR(1) model and with a buy and hold strategy. The neural network of three layers had the best performance and it forecast ability was significant in statistical and economical terms. When we used a recursive process for re-estimate period to period the weights of the network, we do not find a better performance. In addition, we compare the percentage of sign prediction between ward standard network, ward recursive network and the AR(1) model, and the last was significantly higher. However, the model's forecast capacity did not generate higher significant returns, in relation to a passive investment strategy, even when the abnormal average returns was positive. The results were the expected for the Chilean case due to: the few agent that have this market, the high volatility of rate exchange as consequences of international financial crisis, and the significant changes in the market produced by the Central Bank's interventions.
Keywords:
Dollar exchange rate, Neuronal networks model, Chile, Prediction
Parisi F., A., Parisi F., F., & Guerrero, J. L. (2003). Neural network models applied to the prediction of the dollar exchange rate observed in Chile. Estudios De Administración, 10(1), 25–48. https://doi.org/10.5354/0719-0816.2003.56780
