In this study, the hourly movement directions of 9 banking stocks in Borsa Istanbul were predicted using Long-Short Term Memory(LSTM) networks with features obtained from financial news. In the feature creation phase, the word embedding referred as Fasttext, and the financial sentiment dictionary were utilized. Class labels indicating the movement direction were computed based on hourly close prices of the stocks and they were aligned with obtained feature vectors. Two different LSTM networks were trained to perform the prediction, and the performance of the classification process was evaluated by the Macro Averaged (M.A) F-Measure. In the experiments, the movement directions of the 9 stocks were predicted with an average M.A F-measure rate of 0.540. Although the results of both LSTM networks were higher than the Random and Naive benchmark methods, the use of Attention Mechanism in the second LSTM network did not positively affect the results.