基于阶段评价的BP及在股价预测中的应用

BP Based on Phase Evaluation and Its Application in Stock-price Forecasting

研究了BP中收敛速度慢、局部极小值的问题,针对现有研究中的改进方法对BP的性能提升有限,从疲劳的角度出发,提出了一种新的改进方法—阶段化评价法(Phase Eva luation),进行一阶段学习后对性能指标进行评价,并用于调整下一阶段学习方式。改进后的BP网络具有收敛速度快、稳定性强,易于跳出局部极小值的特点。为验证阶段化评价法的有效性,将其用于股票价格的预测与分析,在对股票数据的组织上不是直接拟合股价走势曲线,而是拟合期望收益率的出现概率。实验结果表明基于阶段评价的BP不但学习性能比传统BP高,而且结合文中的股票数据组织方式可以提高实际股票交易的操作性,获得较好的年投资收益率。

The problem of low speed convergence and local minimum in BP and the inapparent upgrade of performance of current reformed BP are studied. From the viewpoint of tiredness, a newly improved method-Phase Evaluation is put forward, which can evaluate the learning performance and regulate the learning mode of next phase. The reformed BP is able to converge faster and more stably, and is able to avoid local minimum more easily. To verify the validity of Phase Evaluation, it is used in the experiment of forecasting and analyzing stock price. In the experiment,fitting the curve of stock price is given up but a new method-fitting the probability of expected income - rate is introduced. The results show that Phase Evaluation has better performance than traditional BP, and can improve operability of stock investment and gain preferable annual - income - rate with the help of fitting probability method.