Crime is the act of offending and doing illegal activity. Previous studies have shown crime should be predicted using demography, location and time features. This thesis attempt to predict crime types using hybrid of deep learning algorithms. 11,363 records (September 2005-April 2011 E.C.) was collected from Bahir Dar city. The data was in paper form and which has been converted to electronic format. After preprocessing the data, we left with 10,270 records. The crime record dataset has been analyzed and tested using the hybrid of feedforward artificial neural network and long short-term memory recurrent neural network. We have compared the performance of FFANN, LSTM-RNN and the hybrid model. MAE, MSE and accuracy is used to measure the model performance. The result for MAE is 0.249, 0.181, and 0.212 for FFANN, LSTM-RNN and the hybrid model respectively. MSE is measured as 0.097, 0.051, and 0.049 for FFANN, LSTM-RNN and the hybrid model respectively. Whereas the accuracy is 90.20%, 94.80%, and 95.07% for FFANN, LSTM-RNN and the hybrid model respectively in which the proposed model outperforms the two algorithms. The study result show that applying the hybrid of FFANN and LSTM-RNN enables to come up with more accurate prediction. Predicting crimes accurately helps to improve crime prevention and to optimize resource allocation in police departments. This crime prediction result can be improved by applying attribute selection by computing information gain and correlation.