Changes in the distribution of financial time series, particularly stock market prices, can happen at a very high frequency. Such changes make the prediction of future behavior very challenging. Application of traditional regression algorithms in this scenario is based on the assumption that all data samples are equally important for model building. Our work examines the use of an alternative data pre-processing approach, whereby knowledge of distribution changes is used to pre-filter the training dataset. Experimental results indicate that this simple and efficient technique can produce effective results and obtain improvements in prediction accuracy when used in conjunction with a range of forecasting techniques.
展开▼
