Exchange Rate Forecasting with Ensemble K-PLS and Ensemble Neural Networks: A case study for the Indian Rupee/US Dollar

The purpose of this presentation is to evaluate and benchmark ensemble methods for time series prediction for daily currency exchange rates using ensemble methods based on feedforward neural networks and kernel partial least squares (K-PLS). Ensemble methods reduce the variance on the forecasts and allow for the assessment of confidence metrics and risk for the forecasting model. The use of neural networks for time series forecasting has been well established and best practice methods are summarized in [1-3]. An obvious advantage of artificial neural networks is that the models are nonlinear and relatively easy to train. Shortcomings of the neural network literature for time series forecasting include (i) the lack of consensus for parameter settings; (ii) a lack of established standards for training neural networks; (iii) a lack of consistent evaluation metrics for time series forecasting; and (iv) a lack of clearly established benchmark problems. It has been shown that averaging neural network forecasts leads to more robust models that furthermore allow for an estimate of the confidence level [4]. Zimmermann reported that typically ensembles of 200 neural networks with the same neural network architecture, but with different seeds for the random weight initialization, are sufficient. Whereas Zimmermann applies ensembles of recurrent neural networks, ensembles of feedforward neural networks trained with the backpropagation algorithm will be applied in this work. In order to let training proceed in an automated fashion an extension of the Efficient BackProp strategy introduced by LeCun was applied [5, 6]. In addition, we propose two different types of ensemble methods: (i) an approach similar to that of Zimmermann, where the different neural networks have the same architecture, but are initialized with different random weights, and (ii) a novel ensemble strategy, where the training models use different weight initializations, but in addition, multiple cross-validation folds are used for training the neural networks. Two different types of time series forecasting methods will be investigated: (i) a one-step ahead prediction, and (ii) a roll-out prediction that will lead to long-term forecasts by feeding predictions back to the input space (just as if they were the actual values) and bootstrapping over successive steps to make multi-step ahead forecasts. A novel ensemble method for time series prediction based on Kernel Partial Least Squares (K-PLS) is also introduced. Kernel partial least squares [7-8] is a " kernelization " of the (linear) Partial Least Squares method. PLS is widely applied in chemometrics and …