A discussion is given on the convergence of the on-line gradient methods for two-layer feedforward neural networks in general cases. The theories are applied to some usual activation functions and energy functions.
Some widely-used technical indexes of stock analysis are introduced as input of BP neural networks for the prediction of ups and downs of stock market, and better accuracy of prediction is achieved. A jump training strategy and three varying training ratio methods are used to accelerate the training iteration. An online prediction strategy is applied to monitor the training iteration procedure. The ratio of central distances of prediction examples is defined, in order to locate the un-stable prediction examples.
