Financial time series are inherently chaotic, posing significant challenges for accurate forecasting. Recently, one effective method for financial time series forecasting is combining Phase Space Reconstruction method with a deep neural network, such as Long-Short-Term-Memory (LSTM). However, most of time series forecasting methods, even the deep learning models, are sensitive to anomalies or outliers in the time series. Anomalies (discords) bring about negative effects on the performance of time series forecasting. There have been some research works which suggested a novel framework to improve time series prediction through anomaly detection-and-repair in a preprocessing step before forecasting. In this study, we attempt to apply this framework to enhance the accuracy of financial time series forecasting. Our proposed method consists of two stages: first, anomaly detection-and-repair is used as a preprocessing step and then Phase Space Reconstruction method combined with Bi-directional LSTM model is used as time series predictor. Our proposed method is called AD_PSR_BiLSTM. Results of experiments on six financial time series datasets reveal that the proposed approach AD_PSR_BiLSTM can improve the prediction accuracy of the simpler forecasting method: Phase-Space Reconstruction combined with BiLSTM (called PSR_BiLSTM).