Business applications make extensive usage of time series analysis for the most diverse tasks. By analyzing the development of any phenomena over time we gain some useful insights on the stock market forecast, analyze the risk related to investments, understand the behavior of a company on the market and so on. More specifically, in a corporate investment banking environment, analyzing the transaction history of a customer over the years is crucial to establish a fruitful relationship and adapt to its behavioural changes. In this environment we recognize three macro-categories of phenomena of interest: cyclic events, sudden and significant changes in trend, and isolated anomalous points. In this paper we present a framework to automatically spot these behaviors by means of simple - yet effective - machine learning techniques.We observe that cyclic behaviors and sudden changes can be easily targeted by means of adaptive threshold algorithms, while unsupervised machine learning techniques are the most reliable in detecting isolated anomalies. We design and test our algorithms on actual transactions collected in the past two years from more than 2,000 customers of UniCredit Bank, showing the efficiency of our solution. This work is tested to serve as a decision aid tool for corporate investment banking employees to facilitate the inspection of years of transactions and ease the visualization of interesting events in the customer history.

Regular pattern and anomaly detection on corporate transaction time series / Soro, F.; Mellia, M.; Russo, N.. - ELETTRONICO. - 2578:(2020). (Intervento presentato al convegno Workshops of the 23rd International Conference on Extending Database Technology/23rd International Conference on Database Theory, EDBT-ICDT-WS 2020 tenutosi a dnk nel 2020).

Regular pattern and anomaly detection on corporate transaction time series

Soro F.;Mellia M.;
2020

Abstract

Business applications make extensive usage of time series analysis for the most diverse tasks. By analyzing the development of any phenomena over time we gain some useful insights on the stock market forecast, analyze the risk related to investments, understand the behavior of a company on the market and so on. More specifically, in a corporate investment banking environment, analyzing the transaction history of a customer over the years is crucial to establish a fruitful relationship and adapt to its behavioural changes. In this environment we recognize three macro-categories of phenomena of interest: cyclic events, sudden and significant changes in trend, and isolated anomalous points. In this paper we present a framework to automatically spot these behaviors by means of simple - yet effective - machine learning techniques.We observe that cyclic behaviors and sudden changes can be easily targeted by means of adaptive threshold algorithms, while unsupervised machine learning techniques are the most reliable in detecting isolated anomalies. We design and test our algorithms on actual transactions collected in the past two years from more than 2,000 customers of UniCredit Bank, showing the efficiency of our solution. This work is tested to serve as a decision aid tool for corporate investment banking employees to facilitate the inspection of years of transactions and ease the visualization of interesting events in the customer history.
2020
File in questo prodotto:
File Dimensione Formato  
DARLIAP7.pdf

accesso aperto

Descrizione: Articolo principale
Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Creative commons
Dimensione 913 kB
Formato Adobe PDF
913 kB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2821832