The Speed of Now: Using AI to Manage Liquidity
Most financial institutions
are acclimating to real-time payments in their daily operations. Even if most
of the participants in the FedNow® Service are set up to only receive payments
for now (rather than receiving and sending), instant payments are here
to stay.
Instant payments are also poised to radically change the way that banks approach liquidity and treasury management.
The current strategies and protocols for maintaining good liquidity are built around the current batch payment ecosystems. However, consumers are eager to enjoy the benefits of instant payments.
Banks and credit unions need to rethink their playbooks around liquidity management. They also need new tools that can handle higher payment volumes in real time. Enter the technological marvel of artificial intelligence and machine learning (AI/ML).
What Makes AI/ML a Good Solution For Liquidity Management?
A quick distinction: machine learning is a subcategory of the broader discipline known as artificial intelligence. Machine learning is focused on building statistical models and algorithms that can process data with less human intervention.
Machine learning can support financial institutions by crunching through petabytes of data and identifying meaningful patterns.
Although machine learning isn’t a new branch of AI, the lack of usable liquidity data within the financial services industry has meant that this type of technology couldn’t offer much help. That’s changed in recent years.
For financial institutions to use machine learning algorithms to support their liquidity strategies, there are two vital considerations to make:
False Correlations
Because a machine learning algorithm isn’t context-aware, it can be prone to making false correlations. For example, a positive correlation emerges if you compare a graph of the Federal Reserve’s benchmark interest rate over time with the physical heights of the Fed Chairpersons. That’s a ludicrous connection to make. But an algorithm can’t discern that for itself.
When implementing a machine learning algorithm to monitor liquidity data, financial institutions should establish checks and balances to flag potential issues and prevent them from swaying important decisions.
Transparency
The number of scientists and computer engineers who understand how machine learning works is small, but it’s orders of magnitude larger than the number of bankers who understand machine learning and the algorithms that drive these models.
However, bankers do carry responsibility for the tools and models they use to manage liquidity and the solvency of the institution overall. Bankers will need transparency into the models, i.e., the reasoning and methodology for making the decision.
The need for transparency has given rise to a phrase: explainable artificial intelligence (XAI). Humans need to be able to trust the algorithms and interpret how they produce a given output.
Drinking From a Firehose of Data Instead of a Drip
Digital payment volume is on an upward trajectory with no end in sight. Combined with the adoption of FedNow and the ISO 20020 data standard, the world of payment data has gone from a drip to a deluge in a very short period.
This is excellent news for institutions ready to adopt machine learning algorithms into their liquidity management strategy. More data allows the models to improve their accuracy and capabilities, which can eliminate concerns like those mentioned above around false correlations.
Here is the basic process for training an AI model:
1. Feed the AI a historical data set.
2. The model learns the data set and identifies patterns and correlations.
3. Feed the AI model fresh data in the format it was trained on.
4. The model can then offer predictions and recommendations on the next best action.
3 Applications for AI/ML in Liquidity Management
With enough data and well-trained machine learning or AI models, three compelling use cases emerge:
Payments Fraud Detection
Real-time payments also mean real-time fraud. That leaves little room for institutions to rely on manual review processes. The good news is that machine learning algorithms can be trained on historical payment data and taught to identify fraudulent transactions. This provides a baseline to identify future anomalous payments that present a higher fraud risk.
Payments Optimization
As institutions build out products and services using real-time payment rails, and in many cases multiple payment rails, they will need tools to manage high volumes of instant payments. This activity also directly affects the bank's liquidity position and what steps should be taken to maintain proper levels.