Identifying Financial Institutions by Transaction Signatures

TL;DR

This paper explores supervised learning techniques, including NLP methods, to identify the financial institution sender of transactions from description strings, using a large dataset of over 10 million transactions.

Contribution

It introduces a novel approach combining traditional and RNN-based NLP methods to infer FI identity from transaction descriptions, addressing normalization and standardization challenges.

Findings

RNN-based methods outperform traditional NLP techniques.

High accuracy achieved on large real-world dataset.

Effective identification of issuing financial institutions.

Abstract

Financial data aggregators and Personal Financial Management (PFM) services are software products that help individuals manage personal finances by collecting information from multiple accounts at various Financial Institutes (FIs), presenting data in a coherent and concentrated way, and highlighting insights and suggestions. Money transfers consist of two sides and a direction. From the perspective of a financial data aggregator, an incoming transaction consists of a date, an amount, and a description string, but not the explicit identity of the sending FI. In this paper we investigate supervised learning based methods to infer the identity of the sending FI from the description string of a money transfer transaction, using a blend of traditional and RNN based NLP methods. Our approach is based on the observation that the textual description field associated with a transactions is subjected to various types of normalizations and standardizations, resulting in unique patterns that identify the issuer. We compare multiple methods using a large real-word dataset of over $10$ million transactions.