Polynomial neural networks with a weighted direct solution and their application in stock index prediction

Abstract

A method for increasing the forecasting accuracy of stock indexes has long been a key problem in financial forecasting and management fields. To resolve this problem, we propose a stock index forecasting model based on generalised multivariate polynomial neural network (PNN). First, a single-hidden-layer generalised multivariate PNN is designed, and an optimal weight vector is proven to exist. Next, the importance value index is creatively designed, and partial derivative analysis is introduced to resolve the issue of neural networks lacking interpretative capabilities. Moreover, we describe the design of a direct solution of weight and confirm the weight vector produced through this solution is the best one. Finally, we build a stock index forecasting model based on generalised multivariate PNN and design a MATLAB-based graphical user interface (GUI). Through empirical analysis and comparison, the effectiveness and usefulness of the above model and its GUI in stock index forecasting is tested.