The Research And Application Of The Online Algorithms

Because of its high noisy, nonlinear and decision factor uncertain features, financial forecasting is an active study field for both investors and scholars. On the basis of the conclusion made of the problems in the applications of financial time series forecasting with neural networks, the nearest neighbor-clustering algorithm is improved. And the algorithm is effective. 1.1 The optimization of width for the nearest neighbor-clustering algorithm Although, the nearest neighbor-clustering algorithm is excellent in shorting training time and less computing quantity, the total simulated effectiveness is not very satisfactory. The reason is that the order of the input samples is important in time series forecasting which is not considered in the nearest neighbor-clustering algorithm. The selection of width has greater effect on its performance. If the width is taken smaller, the last data can be fitted better, whereas if the width is taken larger, the front data can be fitted better. Based on these observations, we change the width. The width is calculated as follows: r=(0.18/total_kk)k+0.01Where total _kk is total of sample. k is current sample. The experiment result shows that this kind of improved algorithm is also excellent in shorting training time and less computing quantity. At the same time, the improved algorithm is excellent in capability of forecasting. 1.2 The improvement of the nearest neighbor-clustering algorithm Although the hidden node can be modified in training process by using the nearest neighbor-clustering algorithm, the hidden node only can be added. For the system of large samples, if the hidden node only is added and not be controlled, the structure of network will very large. The algorithm will create many redundant hidden nodes. So we bring forward a new algorithm. The new algorithm is able to modify the structure of the network (the number of nodes in the hidden layer). (1) Check if a hidden node is deleted We use a matrix H to realize deleting of the redundant node. (2) Calculating of connection weights We use two ways to calculate connection weights. If the structure of network is not changed, we use Recursive least square method (RLS) to adjust weight. If the structure of network is changed, we use Standard least square method to adjust weight. Recursive least square method: 1( ) ( 1) ( 1)( ( ) ( ) ( 1))( ) ( 1) ( ) ( ) ( 1) ( )( ) ( ) ( ) ( 1)TTTw k w k q k y k z k w kq k p k z k z k p k z kp k I q k z k P kλλ???????? === ?? ?? ? + ?? +??+ ?? ???? Standard least square method: ( ) ( ( ) ( )) ( ) ( )w k = R T k ? R k ?1? R Tk ?s k...