Recurrent neural networks have received vast amount of attention in time series prediction due to their flexibility in capturing dependencies on various scales. However, as in most of the classical forecasting methods, its accuracy is strongly tied to the degree of signal complexity. Specifically, stock market prices are commonly classified to be non-linear, non-stationary and chaotic signals, since they exhibit erratic behavior that conducts a poor performance in the long short-term memory (LSTM). In this paper, we propose a methodology to improve the predictability of financial time series by using the complete ensemble empirical mode decomposition with adaptive noise and the intrinsic sample entropy (SampEn). We evaluated the integrated model by applying it to S&P 500 index stocks for the period between January 2018 and April 2020 and for each time series of stock closing prices, an LSTM model was trained to forecast the next closing price. The experimental results represent a dependency between the decomposed signal entropy and the performance of forecast accuracy. This suggests that in those cases where the short-term complexity in financial time series is smaller compared to the series energy, the forecasting capabilities are significantly improved after the removal of decomposed highest frequency. Furthermore, our results show an improvement in forecasting the direction of the stock price by 31% using the classical LSTM architecture.