The final piece of the puzzle has just been published.

First, it was my white paper called “Busting Financial Crime with TIBCO“, which explains the hows and the whys of our end-to-end approach to this horizontal issues. In particular, currently common enterprise crime fighting strategies suffer from two main problems:

—Too many false positives that keep investigators busy on irrelevant alerts and make customers unhappy.
—Long investigation times, as investigators manually consult disparate data sources for information about an alert.

We solve the first by applying machine learning techniques not only to optimize true positive rates but also to incorporate the ability to detect in the data types of fraud that had not been seen before. And we solve the second by improving the real-time response system, not only compiling the entire context of each alert in a visually compelling, unified source but also tracking any investigative actions to ensure future auditability.

Our solution, however, is not a black box that requires expensive consultancy each time you must update your model to the ever-changing fraudsters’ creativity. The system is transparent and puts our customers in complete control with easy and intuitive dashboards. Because it is based on data, you can re-use it on different types of financial crime, from money-laundering to credit card fraud, from insurance claim management to trade surveillance.

Additionally, we published a supporting Spotfire template, which uses both visual analytics and machine learning in very accessible manner. Business users can load their own historic data with their favourite KPIs and use the template to seize the financial crime patterns present in their organisations.

For example, the simple chart in Figure 1 below is deceivingly powerful. It allows you to harness inspiration for which KPIs or rules you want to control. The X-axis should receive the level of aggregation of your data that you want to investigate, be it transactions, customers, branches, employees, or regions. In the Y-axis, you can inscribe any KPI that you feel may elucidate unusual patterns. Here, we use the number of transactions on the same credit card in the last 24 hours. When you find, as in the example below, that the majority of people have a very stable behavior (in our case, around 0) and just a few users have really unusual values, that is something that merits investigation.

Figure 1: Get inspired to find your KPIs

Figure 2 allows business users to invoke a type of machine learning model that specializes in separating 0s from 1s in historic data. You can use past identified crime cases to build a more efficient strategy for finding them in the future. The blue shaded list allows the user to select all the KPIs available in the data, including those built by you. The orange to blue bar chart shows the ability of each KPI in detecting fraud. Its analysis can elucidate areas of improvement in existing internal control systems, areas of poor data quality, or just areas relevant for understanding crime in the organization. Other outputs are created when calling the model, such as intuitive visualizations to control its quality and which specific lines in the new data are more likely to represent crime. This type of model alone is however insufficient, as it so specializes in finding past fraud that it forgets to react to new fraud.

Figure 2: Apply machine learning and navigate your rules

Figure 3 encompasses our strategy for spotting new and surprising types of fraud. It allows characterizing normal transactions and against them spotting unusual ones, even if they have never occurred before. Unusual transactions should be investigated even if dissimilar to past fraud.

Figure 3: Find new and surprising crime patterns

The template also has other abilities, such as what-if analysis for setting thresholds to the results of both models in awareness of the investigative burden they are likely to represent, and sending one or both of the models to the underlying real-time processing engine at the click of a button together with model versioning information and identification of the user who asked for a model update.

Because all calculations that happen under the hood are written in TERR (TIBCO’s R) and embedded into the template, if required they can be altered by your data scientists. For example, you might want your models to run on a Big Data Spark cluster—no problem!

As a third and last point, we just published TIBCO’s Financial Crime Accelerator, which adds to the above the real-time response system—with full documentation—that allows independence to set it up. The Accelerator includes:

—Streambase workflow that receives the models and thresholds from Spotfire, applies the model to the streams of transactions in real time, keeps track of model versioning, creates a new case in TIBCO’s Business Process Management tool for all alerts, and sends emails to investigators. It also sends all output data to Live Datamart and LiveView, which allows visualizing the flow in real time (Figure 4.).

Figure 4: Visualizing alerts in real time

—BPM setup that serves as investigators’ front-end, such that all actions are auditable in the future. The embedded Spotfire not only collates the context of each alert from all number of data sources (Figure 5), but also allows managing the process to identify bottlenecks, spot users who follow inconsistent procedures, track model quality and suggest its revision.

Figure 5: Consult the full context of an alert efficiently

You can now, open-source and free of charge, learn how to create your end-to-end real-time self-learning platform that can be driven by your own business users from easy to use dashboards for fighting different types of financial crime. Check out TIBCO’s Insight Platform and Spotfire analytics tool, and download and enjoy the Financial Crime Busting Accelerator today!

Let’s block ads! (Why?)

The TIBCO Blog