We consider time series representing a wide variety of risk factors in the context of financial risk management. A major issue of these data is the presence of anomalies that induce a miscalibration of the models used to quantify and manage risk, whence potentially erroneous risk measures on their basis. Therefore, the detection of anomalies is of utmost importance in financial risk management. We propose an approach that aims at improving anomaly detection on financial time series, overcoming most of the inherent difficulties. One first concern is to extract from the time series valuable features that ease the anomaly detection task. This step is ensured through a compression and reconstruction of the data with the application of principal component analysis. We define an anomaly score using a feed-forward neural network. A time series is deemed contaminated when its anomaly score exceeds a given cutoff. This cutoff value is not a hand-set parameter, instead it is calibrated as a parameter of the neural network throughout the minimisation of a customized loss function. The efficiency of the proposed model with respect to several well-known anomaly detection algorithms is numerically demonstrated. We show on a practical case of value-at-risk estimation, that the estimation errors are reduced when the proposed anomaly detection model is used, together with a naive imputation approach to correct the anomaly.