Accurate financial time series forecasting is important in financial markets. However, for financial time series with low fluctuation, there is an unusual forecasting phenomenon in the popular recurrent network model forecasting, with the predictive value lagging the truth value. We call this phenomenon the lagging problem. This study proposes new evaluation measures for assessing the lagging problem, including lagging relative error, lagging value error, and lagging trend error. Moreover, the state analysis method and linear fitting model are developed to explain the causes of the lagging problem. Experimental results show that all popular recurrent network models adopted suffer from the lagging problem. This problem is caused by the failure of the nonlinear function in the prediction model and the linear degeneration of the prediction model thereafter, resulting in the suppression of the nonlinear fitting ability.