A comparison study of performance measures and length of intervals in fuzzy time series by neural networks

Abstract

Deciding length of intervals and choosing performance measures have important issues to forecast fuzzy time series. Many forecasting studies accept MSE(Mean squared error) for performance measure and use only one kind of length of intervals such as 1000 without showing any reason and this situation significantly affects forecasting results. This study applies a backpropagation neural network to forecast fuzzy time series with different performance measures and length of intervals. ISE (Istanbul stock exchange) national-100 index for the years 2001-2008 is used for forecasting target. MSE, RMSE(Root mean squared error), MAE(Mean absolute error) and MAPE(Mean absolute percentage error) for performance measures are compared for different length of intervals. The experimental results show that 300 as length of intervals outperforms other lengths of intervals in overall performance of MSE, RMSE, MAE and MAPE for forecasting ISE national-100 index.