A study on a Kalman filter and recursive parameter estimation approach applied to stock prediction.

Abstract

This thesis describes a first experimental project using a recursive parameter estimation and Kalman filter approach to on-line modelling and prediction of stock market time-series. On-line (real-time) and daily closing price stock data are identified as Box-Jenkins ARIMA models. Differencing is performed to obtain a locally wide sense stationary process which is identified through spectral estimation methods. The initial model parameters are updated on-line via the Recursive Prediction Error algorithm and predictions are performed using the Kalman filter. This approach is studied and compared to the traditional Box-Jenkins SISO approach. The daily stock processes are also modeled as autoregressive processes embedded in white noise, which make an ideal investigation for the Kalman filter.