Detecting anomalies in financial time series with the LSTM-AE SageMaker algorithm

Anomaly detection in financial time series plays a crucial role in identifying unusual market conditions that could affect trading strategies and pose risks to investors. Anomaly detection can help identify abnormal price movements or trading volumes associated with specific events, such as earnings announcements, release of economic indicators, or geopolitical tensions. Anomaly detection algorithms are also useful for automatically detecting and correcting data quality issues in the market data time series used for calculating portfolio risk measures [1].

Recurrent autoencoders are standard deep learning architectures for detecting anomalies in sequential data. The autoencoder is trained in an unsupervised manner to learn a low-dimensional latent representation of the data (through the encoder), and to reconstruct the normal behavior of the data from this latent representation (through the decoder). The trained autoencoder is then applied to new unseen data, and an anomaly is detected whenever the observed data deviates significantly from the autoencoder’s reconstruction.

Different autoencoder architectures have been proposed in the literature on time series anomaly detection [2]. In this post, we will focus on the Long Short Term Memory Networks based Encoder-Decoder scheme for…