A method for outlier time series detection based on one-step chaotic prediction
Tao Yeqing · Engineering Journal of Wuhan University · 2010
Traditional outlier detection has been dominated by statistical method to find out those unusual behaviors deviating from statistical profile.We propose a new outlier detection method based on chaotic behavior prediction theory.The method firstly tries to understand the nature of short-term financial transaction operations through time series chaotic analysis,and then provides a mechanism of forecasting the next step result.We construct an approximator for financial operations and a predictor for one-step behavior by employing radial basis function neural network,to detect those behaviors apparently contradicted to the prediction.Experiments on synthetic data mixed with real-world data and simulated outlier data demonstrate encouraging performance in outlier extraction.