Back

Speaker "Leah Mcfarland" Details Back

 

Topic

Challenges in using machine learning for financial monitoring

Abstract

Shifting from rules-based to machine learning-based financial monitoring reduces false positive alerts and increases escalation rates (ie, alerts selected for further investigation)
In conjunction with entity resolution/network generation logic, it also enables enhanced visualizations in support of investigations
However, notwithstanding the above, there continue to be multiple challenges in making this important transition, including:
Supervised machine learning uses suspicious activity reports (SARs) as a target, but these may be of low quality (eg, “defensive” filing)
Changes in underlying data over time (ie, data drift) require model retraining, but regulatory requirements on model validation extend this process
To move toward real-time data processing and/or more continuous model retraining requires greater maturity in model lifecycle management
There remains much greater scope for reduction of false positives, given:
  • Rates are still high even after initial machine learning adoption
  • Historical approaches to monitoring models have been backward (eg, identify rare data pattern and impose blanket rules)
  • Regulations and their application are ambiguous in not significantly distinguishing between low- and high-risk activities
  • There is insufficient attention to how compliance costs serve as a barrier to access to financial services

Increasing adoption of blockchain involves a novel data processing/storage mechanism only beginning to be subjected to financial monitoring

There remains a perception that machine learning is “black box,” despite relying on well understood statistical techniques

Use of machine learning has been associated with perpetuation of gender, racial, and other biases (and for financial monitoring could be relevant for treatment of money services businesses)
Development of machine learning models depends on availability of data scientists, who are in relatively short supply and have high rates of staff turnover
There are continuing and new data-related issues:
  • Poor data quality from legacy/siloed systems persists within a given institution
  • Third-party proprietary and open source data need to be integrated for holistic understanding
  • Semi- and unstructured data should be leveraged going forward
  • There is growing expectation that data be evaluated in nearer to real-time
  • Cross-border data requirements are becoming stricter

Profile

Leah McFarland has worked as a Director of Technical Program and Product Management at major U.S. banks, including JPMorgan Chase, State Street, and Citibank, where she has focused on data management, cloud migration, and the use of AI/ML and automation, with a specialty in financial monitoring issues.
 
Prior to that, Leah worked at FSVC, where she managed a $5 million portfolio delivering risk management assistance to financial institutions in the Middle East, North Africa, Sub-Saharan Africa, and Eastern Europe.
 
She began her career as a diplomat for the U.S. Department of State, serving in Moscow, the Secretary of State’s operations center in Washington, DC, and Sao Paulo.
 
Leah has an M.A. in International Affairs from the University of California at San Diego and speaks Spanish, Portuguese, Russian, French, and Mandarin.