Intraday stock price analysis and prediction

Abstract

The close relationship between daily and intraday stock data is established using theoretical interpretation and variance estimation by neural network. Based on this, conventional time-series and neural networks are used to analyze the more informative intraday data for stock price prediction. Each method is tried with different set of parameters, in order to obtain an objective and thorough evaluation. The evaluation results show that Widrow-Hoffs LMS should be used given adequate computing resources and time. Back Propagation is optimal if the input parameters of the series are precisely known. ARMAX is a simple and parameter insensitive method. In general, it is a bad choice to use the trading volume as an exogenous input. Contradicting intuition, simple models give better predictions than complex ones, and lightly trained is better than heavily trained.