Algorithms For Anomaly Detection on Time Series: A Use Case on Banking Data

Abstract

The present research aims to present an overview of methods for automatically detecting anomalies in data representing time series. A time series is a sequence of qualitative values obtained at successive times, generally measured with equal intervals. Time series can represent different real-life phenomena, such as the behaviour of the stock market, variations in temperature and other meteorological data, the behaviour of banking credit/debit card consumption, among others. In addition, this work presents a 4-step methodology for preprocessing data and detecting anomalies on a time series dataset representing the spending of debit and credit card customers. A synthetic anomaly injection technique was applied to validate the models. Results can be used to monitor banking behaviour and trigger alarms in case of possible fraud or rare events.