Stock price prediction using intraday and AHIPMI data

Partitioned linear-nonlinear models are developed to improve in-sample precision and reduce sensitivity to out-sample modelling errors in stock price predictions. Such partitioned models are compared with linear regression models and nonlinear neural network models. The partitioned models demonstrate similar performance to the nonlinear models in both in-sample and out-sample predictions. Robust prediction schemes are then introduced to improve the predictabilities of partitioned models. Such partitioned models with robust schemes outperformed both linear regression models and nonlinear neural networks models in terms of prediction accuracy as well as model robustness. In addition, a linear relationship of non-model-based correlation and linear-regression-model-based predictability is found to exist between intraday (as well as AHI-PMI) data and stock price indexes of open, close, high and low.