Turn frequent anomalies into rare exceptions

Investment managers aspire to be data-driven, but the reality is that system fragmentation, operational risk and limited transparency prevent firms from fully leveraging all of their data.

Fragmented research, order management, trade execution and portfolio management systems often operate in silos, resulting in too many systems without a unified view or governance standards. Manual processes and workarounds increase operating risks, exposing you to sub-optimal performance, inaccurate reporting and potential regulatory fines.

Together with limited transparency into the accuracy, timeliness and consistency of data, investment managers are struggling to make decisions based on stale, suboptimal information.

Enter the era of artificial intelligence (AI) and the potential use-case for data validation. AI is well suited to the many challenges of ensuring consistent, high-quality investment data because it identifies potential anomalies across data domains and continuously learns and evolves its understanding as it ingests new data. AI is intuitive: It considers each data domain, looking for anomalies in FX rates, securities prices and transactions.

As AI is applied across the investment process, its understanding of data, true anomalies and false alerts evolves. It’s this feedback loop that enhances the model, creating continuous improvement optimized for each data domain.

At State Street, our team of data scientists and engineers have built several deep learning models, first starting with the market value of a portfolio. These models consider various factors and determine whether there's a problem with the client’s portfolio market value that requires escalation to an analyst for verification and investigation. AI-based anomaly detection replaces traditional, static-based rules, and works by learning and recognizing patterns in the portfolio.

While the previous static-based rules flagged over 31,000 data exceptions across a six-month period (of which there were only 250 true exceptions), the AI model identified only 4,000 exceptions while discovering 100 percent of the true exceptions. The massive reduction in false positives means a 25x productivity gain for the data operations teams.

Security reference data is the next domain where we’re applying AI for anomaly detection. The model determines if there is a suspicious reference data element that is out of sync with other attributes of that security, then escalates for further investigation to a data operations specialist.

We started with market value and security reference data because they cover the most territory in terms of error detection and mitigation. Almost every data error impacts market value in some way, whether that’s due to a corporate action that was improperly processed or a mispriced security in the portfolio.

Looking to the future, we’re expanding the models into new data domains like intraday transactions. If we can identify potential problems with a transaction in real time, we have an opportunity to fix it before end-of-day when the impact is greater. 

We built the models to be source agnostic, providing the ability to run a broad range of different datasets through the process, whether that data originates from our asset servicing operations, gets acquired from third-party vendors or is generated internally by investment firms. Leveraging this transformative technology enables you to streamline your internal data operations and benefit from consistent, accurate data across your investment process.