Facing the nonlinear, dynamic and time-series characteristics of financial markets, this study adopts integrated classification algorithm and long-term and short-term memory (LSTM) model to study the accuracy of financial forecasting. Firstly, the study collects financial data through government regulators, stock exchanges, trade associations and financial data service providers, and then preprocesses the data. On this basis, an integrated classification model based on Support Vector Machine (SVM), K Nearest Neighbor (KNN) and Naive Bayes is constructed, and random forest is used as a meta-learner. The prediction results of several algorithms are combined by stack integration method to improve the accuracy of financial credit default prediction. The study also adopts LSTM model to process time series data and uses its memory function to capture long-term dependence to further improve the accuracy of prediction. In the model training stage, cross-validation and other technologies are used to train and optimize the parameters of the integrated classification algorithm and LSTM model. The results show that the risk prediction accuracy of the integrated classification model in financial data reaches 94.84% on average, which is higher than that of the LSTM model. Based on the above research, this paper finally draws the conclusion that the combination of ensemble learning and LSTM can improve the accuracy of financial forecasting.