Correlation-predictability analysis for intraday predictions

Intraday financial data can be interpreted based on high-frequency and low-frequency time-series modeling. Recent study has revealed the complexity of high-frequency dynamics using correlation analysis, in related with randomized matrices, eigen-decomposition, and hierarchical grouping of stocks. As an alternative approach, we present some ideas on low-frequency "news" modeling as applied to intraday data. A comparative eigen-analysis is described, showing the regularity of correlation matrix and the significance of variance-weighted principal components. It is also shown that low-frequency modeling is related with a receding-horizon intraday prediction problem, where improved predictability is conditional upon the available information up to the current time. Strong empirical evidence is obtained for a linear correlation-predictability relationship for intraday high, low and close prediction, which shows promises in applying to the NASDAQ composite index and other financial data. The relationship implies that a non-model based correlation measure can predict the performance of linear regression prediction model that uses intraday information.