Proposal of Qualitative Modeling for Time Series Phenomena based on the Transition of Time Series Data Pattern
Hidenori Naoe · IEEJ Transactions on Electronics Information and Systems · 1999
The present paper proposes a modeling method which describes and predicts time series phenomena in a qualitative manner. For the qualitative description, time series data are characterized by the first and second order coefficients which are the outputs of a time dependent polynomial filter. In order to predict by the neural net, the coefficients are transformed and normalized to [0, 1] space. From the series of these normalized coefficoients, a few teaching patterns for neural net are extracted by clustering algorothm. And by error back propagation algorithm, the future state of the time series phenomena is obtained as symbols. Because the stock price does not necessarily walk randomly, the stock prices are tried to identify and predict as the time series data. Comparing the proposal method with convential autoregressive model, it is clarified that the former has even high precision concerning with the predicted pattern by the correlation coefficients.