We pexamine the problem of modelling financial time series contaimianted with outliers, trends, and level shifts. The problem is two-fold, preictions based on contaminated data are suspect and models estimated from such aata are distorted. A robust estimation algorithm based on a filter which "cleans" a time series of outliers and adjusts for trends and level shifts is developed. The cleaned and adjusted data is then used to estimate models and as a basis for further predictions. Both linear and nonlinear neural network model sare examined. The performance of the robust algorithm is examined on both an econometric problem of predicting tobacco sales and a finanical problem of modelling the FTSE and the S&P.
展开▼
