Risk Management Accelerator

Introduction

The Risk Management Accelerator illustrates a reference architecture to implement Streaming Machine Learning. The definition of risk in this case is cross industry: the risk that a lot is maverick (manufacturing), the risk that a customer will churn (horizontal), the risk that a transaction is criminal (horizontal), the risk that a set of sensor readings herald a machine failure (manufacturing), the risk that a customer will say no (or yes) to an offering. in general, the Accelerator is relevant whenever we want to use machine learning to single out a subset of our data from the rest. The Accelerator includes a demonstration that shows an Anti-Money Laundering (AML) use case to detect potentially criminal transactions in a data stream.
 

This video shows a walk-through of the Accelerator in action.

This video explains why, within the financial crime use-case, the Accelerator is important and how it works.

Download

The Risk Management Accelerator can be downloaded from the Exchange.

Business Scenario

Companies have always had to make decisions. In an almost remote past, talent was enough to give a company competitive advantage - companies did best which employees had more talent to feel the market and formulate strategies accordingly. In the digital economy however this is increasingly insufficiently, not only because talent has always been scarce but because the availability of data allows us to mathematically support our opinions with evidence. And this again and again beats just talent.
 

Machine Learning is a branch of Artificial Intelligence that allows computers to detect patterns in data and automatically adapt those patterns change. When leveraging Machine Learning in risk scenarios, businesses can more efficiently spot where the riskiest data is today  (e.g. those customers who are more likely to churn or those transactions that are more likely to be credit card fraud), but also understand the likely drivers of that risk.  The net result allows them to draw policies for combating the risky scenarios effectively, e.g. tailored customer messaging that turns customers around for addressing the reasons for churn, or a revision of internal control procedures that address precisely the reason why crime almost went undetected.  In turn, this maximizes customer satisfaction and retention as well as boosts business efficiency and productivity.
 

But machine learning tends to be considered complex by most business users, which inhibits its application across the breadth of organizations. The Accelerator allows a business field expert to get to a machine learning model within ten clicks, evaluate it, understand its drivers and apply it to today's data. The reusability of the platform across use-cases and the transparency of the end-to-end solution frees businesses from consultant dependency, effectively empowering the business to own its own success drivers. The front-end was created as a tutorial that demystifies the complexity and speaks to business experts who understand their KPIs.
 

Today's dynamic risk detection systems need to be agile, scalable and intelligent.

• Agile to adapt to ever changing realities.
• Scalable to deal with ever increasing data volumes.
• Intelligent to detect increasingly sophisticated patterns, and also to reduce false positives in the alerting stage.

Concepts

Risk detection spans many different domains. Within  financial services (FSI), some of these include Anti-money laundering (AML), Trade surveillance, Credit/Debit card monitoring, Health insurance fraud, Insurance fraud, Online operations. In manufacturing, risk can mean risk of producing defective items, or risk of machines working inefficiently, or risk of machines failing. In transport & logistics, there is the risk of losing luggage, or the risk of a car or boat needing repair or being deviated from its optimal course. In customer relations there is the risk of a customer feeling spammed.
 

In all cases what they have in common is a relatively high importance of transactions, at which any given time a certain percentage can be classified as risky and therefore deserving of special treatment. This percentage will vary greatly by market, industry, value, and other parameters. But, even if the content of the data differs greatly among them, the machine learning techniques can, at least initially, be widely similar. 
 

The key to Streaming Machine Learning action systems is to maximize the number of cases that are identified, while minimizing the number of false detections (false positives) to come up with an optimal system both from an operations and financial standpoint.

Benefits and Business Value

The Accelerator brings the following key benefits.
 

On the Predictive Analytics side

• Use your historical data to find patterns that represent risk, 

• Use TERR to create models that capture those patterns, 

• Do what-if analysis to set thresholds for model results so as to adjust your actions to the size of your team or budget,

• Have your business users own this process from easy to use interfaces, without expensive consultancy,

• Data discovery, machine learning model creation and real time deployment within a single tool.

On the Streaming Machine Learning side

• Use the latest best model defined by your business people to

• Score every transaction in real time,

• Keep model versioning,

• Accelerate the adoption and rolling out of real time action: A graphical flow based development reduces complexity and increases collaboration; projects can span multiple assets, data and scenarios; with connectivity to over 150 options, including Bloomberg, BM&FBovespa, Currenex, EBS, FIX, FXall, Hotspot, Interactive Data, and Thomson Reuters, all your data feeds can be included.

Technical Scenario

The Accelerator include a demonstration of fraud detection using a credit card transaction dataset that is analyzed against both a supervised and an unsupervised machine learning model to produce a score that indicates the risk's probability, and another score that show how the transaction deviates from what could be considered normal.
 

Other demo scenarios will vary from this. For example trade surveillance will differ greatly from insurance claims and credit card transactions. However the Accelerator is intended to show a design pattern for detecting fraud using a combination of event processing and analytics rather than provide a series of concrete rules for doing the detection. The agility of the analytics and event processing platforms allow customers to modify the framework quickly and easily to adapt to other scenarios.
 

Components