Transformer based anomaly detection on multivariate time series subledger data

Subledgers maintain detailed information about specific accounts or transactions in order to substantiate the general ledger. Subledgers provide a granular level of detail for financial reporting and analysis, which is especially essential for accounts receivables and payables. The size of subledgers can vary greatly depending on the complexity and volume of transactions and their size can also increase over time as more transactions are recorded. Depending on the size of a company, subledgers may record hundreds of billions of financially significant business events each year originating from its different legal entities. As many accounting customers rely on the subledger as the source of financial results, it is crucial to identify anomalies early on to minimize their impact and prevent further harm. To this end, we present a novel algorithm designed to analyze subledger transactions in a systematic manner. Our approach employs a modified transformer architecture for the task of anomaly detection in multivariate time-series data which relies on its ability to reconstruct input. Moreover, it utilizes the reconstruction loss as a priority value to emphasize data points that may be anomalous. We also enhance the anomaly score by incorporating seasonality and relationships between different attributes of the subledger. Experimental results demonstrate the efficacy of the proposed approach for different types of anomalies.