The Use of Domain Knowledge in Feature Construction for Financial Time Series Prediction

Most of the existing data mining approaches to time series prediction use data preparation techniques involving an embed of the most recent values of the time series, following the traditional linear auto-regressive methodologies. However, in many time series prediction tasks the alternative approach that uses derivative features constructed from the raw data with the help of domain theories can produce significant prediction improvements. This is particularly noticeable when the available data includes multivariate information but the aim is still the prediction of one particular time series, a situation that occurs frequently in financial time series prediction. This paper presents a method of feature construction based on domain knowledge that uses multivariate time series information and improves the accuracy of next-day stock quotes prediction, when compared with the traditional embed of historical values extracted from the original data.