Exchange rate volatility prediction by an extended self-organizing mixture model
一种自组织混合模型在汇率波动性预测中的应用

Exchange rate volatility prediction has long been the issue that most financial market researchers are concerned with. This article applies a self-organizing-map-based method to the self-organizing-mixture model(SOMAR) for predicting the non-stationary volatility of daily exchange rate. This extended SOMAR(ESOMAR) model is free from the constraint of stationarity which is required by most of the traditional regressive models; it also replaces the global modeling by the local modeling by splitting a non-stationary time series into piece-wise stationary time series episodes. Meanwhile, ESOMAR is a non-parametric neural network regressive model; it combines the simplicity of the traditional regressive model and the flexibility of neural networks, making it adaptive to the heterogeneous data. The prediction results of exchange rate volatility show that the ESOMAR outperforms many traditional regressive models as well as other neural-network-based approaches.