A major concern when dealing with financial time series containing various market risk factors is the presence of anomalies. These can induce misadjustment of models used to quantify and manage risk, leading to erroneous risk measurements. We propose an approach that overcomes most of the inherent difficulties and aims to improve anomaly detection in financial time series. Valuable features are extracted from the time series by compressing and reconstructing the data by principal component analysis. Next, define anomaly scores using a feedforward neural network. A time series is considered contaminated when the anomaly score exceeds a given cutoff value. This cutoff value is not a manually set parameter, but tuned as a neural network parameter through customized loss function minimization. The efficiency of the proposed approach compared with several well-known anomaly detection algorithms has been numerically demonstrated on both synthetic and real datasets, showing high and stable performance with the PCA NN approach. Achieved. We show that correcting anomalies using the proposed anomaly detection model with a basic imputation approach reduces the value-at-risk estimation error.