Marwala EMPIRICAL EVALUATION: SIMULATED DATA SET Case.Study.1: Firstly the algorithm proposed above is evaluated on the time series data produced by numerical simulation. A sequence of un-correlated Gaussian random variables is generated with zero mean and variance of 0.108 as prepared by Stephanos and Anthanassoulis (2001). In this chapter, data are simulated as if they are coming from a sensor that measures some variable that exhibit non-stationary characteristics. The data are made to show some cyclic behavior that simulates a cyclic concept drift. Figure 13.1(A) shows a sample of the simulated data. Case.Study.2: The second test data is created using the Dow Jones stock market data. The stock market is well known for being difficult to predict as it exhibits non-stationarity. The opening price of the Dow Jones stock is also simulated as some data collected from some sensor and sampled at a constant interval. Samples of these data are shown in Figure 13.1(B). The relative performance of the algorithm is measured by how close the prediction is to the actual data. The results obtained with this data are summarized in the next section. EXPERIMENTAL RESULTS Firstly, the effect of the number of regressors on the error is evaluated. Mean Square Error is computed for each prediction and results are presented in Figure 13.2. Performance in terms of accuracy is shown