Time series forecast and anomaly detection at scale applied to business metrics in an ERP environment
André Gil Cardeira Martins · Repositório do ISCTE-IUL · 2019
In the business world, dashboards are a widely used analytical mechanism that helps in the decision-making process by displaying insights, key performance indicators, and business metrics. The information provided by this type of mechanism is strongly aggregated, to obtain a high level of summarization and consequently make reading easier. However, the necessary summarization causes “blind spots” to appear by hiding important information such as a sharp drop in revenue from a specific customer, seller, or product/ service. These “blind spots” make it difficult to detect potential business problems and opportunities, which depend on lengthy and thorough additional exploration. Also, the digital transformation process has resulted in a substantial increase in the number of metrics for all systems supporting the business that need to be tracked. Thus, it will be possible to anticipate actions based on the prediction of future behavior, as well as to detect any isolated or successive deviation from the expected behavior. With this dissertation, we intend to promote the acquisition of knowledge from business data through the application of Machine Learning techniques. Based on the Data-Driven Decision-Making process, we intend to propose integration into an ERP application of a mechanism to predict time-series behavior, as well as detecting and measuring possible anomalies. For dealing with a wide diversity of time series, we propose a meta-learning forecasting method that uses a classifier to identify the best forecasting method for each time series. We also propose a new intelligent metric that allows us to sort time series by the accumulated anomaly. The knowledge generated will complement the information provided by the analytical mechanisms typically present in an ERP application (including dashboards). In this way, we intend to contribute to the maximization of profits and reduction of the possibility of error or fraud, as well as waste and consequently mitigate uncertainty and reduce operational risk. Our solution should promote the need to use Machine Learning in Small and Medium Enterprises, and consequently, future implementation of AI-Driven Decision Making. AI-Driven Decision-Making purposes an assertive and automated reaction to problems or opportunities encountered, but whose study is outside the scope of this dissertation.