| Abstract [eng] |
In this work, stock market data is modelled using linear regression and multilayer perceptron models. Using these models, stock market prices are predicted. For each stock multilayer perceptron with di\text{f}ferent activation functions, size of the hidden and input layer are trained. Experiments show that, no model predicts best for all of the cases. Prediction results are associated with nonlinear analysis measures, such as largest Lyapunov exponent and correlation dimension. Automatic systems for computing these measures are presented. The systems estimation accuracy is then checked with known chaotic maps and equations. Moreover, systems are validated with noisy time series. Lastly, time series classification based on nonlinear measures is introduced. Classification is then used with stock market data. Only classification using largest Lyapunov exponent gives promising results, this is due to errors in computing correlation dimension. Results show that this type of classification could be used for filtering out stocks that are harder to predict. |
