As financial institutions seek to adapt to changing patterns of financial crime more quickly, they are reshaping their anti-financial crime teams. Earlier, anti-financial crime teams consisted solely of analysts and investigators, supported by technology. In the last few years, however, it has become much more common to find analytics experts on these teams. This was, in part, precipitated by new model risk management regulations (for example, OCC SR11-7 in the United States). These new regulations describe elements—such as the need for independent review and model validations—of a sound, effective model risk management program.
Initially, members of enterprise modeling teams were loaned for financial crime analytics. However, due to greater focus and a need for more subject matter expertise, anti-financial crime programs began to form their dedicated analytics teams. Lately, as the cost of compliance has increased, the increased need to use machine learning and advanced analytics has led to much larger analytics groups. These analytics groups often have a team of data scientists dedicated to fighting financial crime.
Typically, the primary responsibilities of analytics (modeling) teams embedded within financial crime programs are pre-production scenario tuning, ongoing above-the-line (ATL) tuning, and below-the-line (BTL) tuning. Here, I explore some of the challenges analytics teams face and present some solutions.
Current Approach: Expensive and Slow to Adapt to New Threats
Given the reputational implications associated with the accuracy and effectiveness of models, issues concerning discovery, model validation, and documentation are of obvious concern to the industry. Erroneous or misspecified models may lead to expensive look-backs or regulatory fines. Critical challenges which can impact the productivity and effectiveness of analytics programs include:
- Desktop tools: Many institutions use desktop tools, which cannot scale or integrate with other internal systems. Most importantly, desktop systems do not provide must-have audit trails or data lineage capabilities.
- Limited coding languages: Typical systems provide the ability to code in specific languages only. However, based on the use cases, data volume, and the institution’s existing infrastructure, analytics teams should be able to design models in any available coding language such as R, Python, Scala, SQL.
- Disparate systems: Often, large institutions use multiple systems for different use cases such as authoring new scenarios, scenario tuning, and machine learning. The systems, most of the time, use different coding languages. This requires additional, sometimes manual integration.
- Cumbersome data provisioning: Mostly, analytics systems are not well integrated with existing production systems. Hence, acquiring production data requires a cumbersome data provisioning process. Based on the institution’s policies, it can take days, sometimes weeks, for data provisioning. In many cases, by the time data is provided, it is outdated.
- Lack of intuitive visualizations: Lack of visualization options leaves the analytics team with a limited ability to understand the underlying data, model output, and investigation outcome.
- Limited discovery capabilities: Generally, the risk of discovery is limited to scenario threshold tuning only. Such an approach does not provide the ability to discover completely unknown suspicious behaviors that institutions might be exposed to.
- Lack of governance and documentation: With model risk management regulations, institutions are required to maintain full audit trails and documentation of the modeling process, which is generally unavailable in typical modeling tools. This leads to the maintenance of two different systems—one for modeling and another one for model documentation— which creates more work for the analytics team.
- Lack of explainability: Typical, machine learning tools allow the data scientist to write machine learning models. However, due to a lack of governance and a full audit trail, such projects remain hard for most people to understand.
- Inefficient model deployment process: Once developed, the operationalization of models, especially machine learning models, is a considerable challenge. One of the primary reasons could be the dis-alignment of data attributes between analytics, transaction monitoring, and case management systems.
- Expensive: Overall, due to multiple systems that require integration, the maintenance of model documentation and the operationalization of models makes most analytics initiatives costly.
Getting Ahead of the Curve
Effective criminal pattern discovery and detection require the application of a variety of techniques. An integrated platform that provides a single, unified workbench for graph analytics, data visualization, machine learning, scenario authoring, and testing for financial crime data is key for an efficient, strategic program. Below are five aspects of the analytics program which can be game-changers in handling the challenges I discussed above.
- Centralized data lake: Institutions realize that effective discovery requires that all the financial institution’s transactions, accounts, cases, and other financial crime-related data be brought into an analytical data lake. Engineered as a portal into the enterprise’s financial crimes data lake, a platform that allows automatic data loading significantly reduces time and effort that data scientists spend in preparing data for analysis.
Additionally, because institutions may have data in different types of systems (Oracle database, big data, structured/unstructured, files), the flexibility to connect these is critical. Lastly, the connection strategy should be flexible to different coding languages. For example, if a user prefers to code in Python leveraging data from Hadoop, it should not require a different system. The ability to read and merge data from various sources for a use case (notebook) allows analytics teams the flexibility to access data quickly. This also significantly improves the efficiency of the IT team and will enable them to focus on the actual analysis of data.
- Polyglot scenario authoring: Apache Zeppelin and Jupyter notebooks are the de facto standard development tools for data scientists. Apache Spark is the most prevalent analytics engine for big data. A strategic platform should be able to leverage these open technologies and standards, thereby minimizing the need to re-train data scientists. Furthermore, the ability to use popular data science languages such as R, Python, SQL, and Scala increases modeler productivity. This is possible by leveraging the interpreter concept. An interpreter is a program that directly reads and executes the instructions written in a programming or scripting language without previously compiling the high-level language code into a machine language program. An analytics platform should allow for the configuration of various types of interpreters such as JDBC, R, Pyspark, Python, Spark-Scala, Spark-SQL, and graph query.
- Intuitive visualization: The configuration of comprehensive use cases won’t be effective without proper visualization tools. The ability to visualize the output of the configured code and the flexibility to tailor meaningful outputs will allow analytics teams to understand the output quickly and more effectively.
Graph analytics is proven to be revolutionary, allowing succinct expression of complex money movement patterns, detecting multi-hop relationships, and identifying hubs and spokes of suspicious activities. Data scientists and analysts should be able to interactively explore financial crime data and gain insights into new and emerging financial crime patterns and trends. The ability to analyze historical data available in the database, and forecast the generated patterns using various interpreters is vital.
- Collaboration and governance: In most cases, analytics teams use one system for the core technical activities and a second system for documentation. A single platform that includes full the audit trail capabilities of notebook authoring, approval, publishing, and deployment takes away the additional documentation tasks from the users. Such a platform is more efficient from a model risk and governance regulation perspective.
- Unified platform: Typically, case management and modeling are handled in two separate systems. This means the deployment of new models and the feedback loop from case management to analytics systems is generally not very efficient, and many require manual integration. Continuous discovery and modeling of new criminal behavioral patterns, coupled with the ability to deploy these models rapidly, is a critical requirement in today’s age of continually changing criminal behavior patterns. Graph analytics and machine learning are essential methods for pattern discovery. They are most effective when applied over the data in a comprehensively designed enterprise-wide financial crimes data lake. Therefore, a strategic platform that is designed from the ground up to enable data scientists to discover and model financial crime patterns rapidly will increase data scientist productivity drastically.
With digitalization paving the way for increased customer expectations from their financial services providers, there is a variety of new products and services that are being offered by banks to cater to this demand. This also means that the volume of transactions is multifold when compared to ten years ago. This, on top of drastically changing risk profiles and criminal behavior patterns, puts anti-financial crime initiatives in excessive focus. The involvement of analytics in financial crime programs has become a bare necessity. Hence it is critical to enable data scientists with the right tools that allow them to nip the bud before it is too late. Any missed financial crime always comes with the risk of huge penalties and even higher reputational risks to the banks.
Message me to learn more or join the conversation by commenting below.
For more information, please visit:
Oracle Financial Crime and Compliance Management Solutions: Webpage
Oracle Financial Services: Homepage
Subscribe to our Blogs:
Oracle Financial Services Blogs: Subscription