November 23, 2020 | Written by: Eliza Salkeld
Categorized: AI | Banking | Cloud | FinTech | IBM RegTech Innovations
Share this post:
With analytic techniques such as business rules, statistical models, and machine learning it can be difficult to understand the role of each approach in identifying fraud.
Analytical techniques used in identifying fraud
There are a variety of techniques that are used to detect fraud, many of which fall under the umbrella of business rules, statistical models, and machine learning. Each of these three main approaches have a place in identifying instances of suspicious behavior. Traditionally, fraud detection has used business rule techniques to achieve beneficial outcomes. The set of business rules an organization defines supports the codification of many different requirements. These requirements, which enable verification, are imposed by a specific policy or learned from hard-earned experience. Statistical models in contrast, reveal the relationships between variables, features or other indicators and are commonly used to understand the distributions of an organization’s data. Finally, machine learning models have many commonalities with statistical models; however, their main objective is to generate predictions.
Each of these techniques provide actionable insight into the characteristics of a claim and the associated risk of fraud, waste, or abuse. The Financial Crimes Insight for Claims Fraud offering specializes in helping insurance companies identify these areas of risk and suspicious patterns. This article explores the themes that define each of the three critical techniques in the analysis of insurance claims data.
The first analytic branch utilized in Financial Crimes Insight (FCI) is business rules. Business rules have been a mainstay in the insurance domain. They have been applied in areas such as underwriting, policy enforcement, and fraud detection. A system of business rules is extremely efficient in its use of known indicators to determine fraud and risk. This type of deterministic rule can be used to make an assessment of a given dataset, identifying when a pattern or data point emulates a behavior known to be suspicious. The results generated through business rules can also be incorporated to generate features that are then passed as input to downstream statistical or machine learning models. The use of both approaches creates features based on proven indicators of fraud, and enables the ability to override a predicted outcome by critical key rules. Having a set of known business rules in place adds a layer of complexity and nuance to the predictions generated by more complicated models.
The second branch found in FCI’s analytics process is statistical models. Statistical models are a special class of mathematical models that are considered non-deterministic. Many statistical models draw inferences regarding the data based on the distribution of the information in various fields. Statistical models provide insight into the data and shed light on what end user decisions can be made. These decisions help identify the fields that used for feature generation, group inference, and even business rule ranges. Another branch of statistical models uses the aggregation of data across multiple dimensions and the ability to determine ratios of certain groups within a population. These models provide an understanding of the distribution of the data, which then aids in techniques such as anomaly detection. FCI for Claims Fraud employs statistical models in both ways enriching both the classification process and generating statistical aggregations across specific dimensions.
The final analytic discussed here is machine learning. While drawing on similar underlying principals found in statistical models, machine learning models instead focus on producing predictions. These predictions are based on the analysis of known outcomes, often referred to as “ground truth.” Given a historical dataset of insurance claim information and related behaviors, the ground truth provides a label that reveals the outcome of each claim. While the type of outcome can vary between insurance companies, it often includes the options such as a “valid” claim or a “fraudulent” claim. Machine learning models will consider the input data from the claim information, indicators produced from the business rules, and features generated from the statistical models. Through this process machine learning models work to identify what patterns are most connected to or telling of a “valid” claim, and what alternative patterns most often lead to a “fraudulent” claim. The models can identify complex and sometimes nonlinear patterns across an expansive amount of data that a human would not likely be able to recognize. Once the most relevant patterns are learned, the machine learning models can be applied to new insurance claim information to predict whether it is most likely to be “valid” or “fraudulent.” Integrating this classification step into the analytics process allows for more intelligent and informed decisions to be made by the end user.
Specific types of machine learning models, such as neural networks, natural language processing, and network graph analytics are discussed in an upcoming entry.
Each of the three overarching techniques described above has specific strengths, and sometimes weaknesses, based on the data available. The Financial Crimes Insight for Claims Fraud product implements each option based on its specific capabilities and known stakeholder needs. While one approach – whether it be rules, statistical models, or machine learning – provides valuable information on its own, it is the combination of all three that creates a powerful and actionable set of insights. By leveraging an adaptive ensemble model, the results, understandings, and predictions can be synthesized to accommodate any blindspots and compound on known strengths. This final analytic result is one that is created from a diverse set of techniques and provides layers of texture and discernment in the information presented to an end-user.
IBM Financial Crimes Insight for Claims Fraud is part of the IBM RegTech regulatory compliance solutions that are designed to help financial institutions better meet their regulatory monitoring, reporting, compliance and risk management needs.
Learn more about IBM Claims Fraud solutions at Financial Crimes Insight for Claims Fraud