Optimizing FX STP through intelligent automation

Today’s digital and data-driven economy has placed an ever-increasing range of challenges on FX market participants. Technological advancements make it imperative that organizations re-think traditional business models, particularly addressing their digital performance and capabilities.

Customers’ demands have evolved to expect faster, more personalized, and reliable digital services. In addition, the COVID-19 pandemic and its fallout have accelerated global digital transformation efforts. Despite this shift, FX operations within these strategies are often overlooked. FX operations must be seen as a crucial enabler of innovative digital products and services. Once optimized, FX operations can enable organizations to meet customer demands and expectations, by allowing highly automated, real-time, and insights-driven operating models.

As Artificial Intelligence (AI) and Machine Learning (ML) methods continue to develop, various rapidly evolving solutions have the potential to add intelligence to automation and take FX operations to unprecedented levels of efficiency. In addition to boosting straight-through processing (STP) rates, these methods can provide actionable insights that increase clarity and control, while catering to customer interactions and alleviating regulatory concerns. The democratization and adoption of AI and ML will continue to drive best practices, leveling the playing field across FX market participants.

Straight-through process(STP)

The FX industry has made significant strides with respect to automation of operations, often achieving relatively high rates of STP within specific segments and corridors. Most notably, around wholesale FX and where CLS is employed, ensuring settlement finality.

As FX operations have become more complex, market participants today face an increasingly diverse set of market forces, such as a new competitive landscape through the emergence of challenger banks and non-bank FinTechs, increased FX market fragmentation, trading volumes, and complexity of products, necessitating greater operational processing capacity. Ongoing regulatory and compliance changes continue to drive new workflows and require conformance both now and in the future.

Furthermore, the transformation of cross-border payments, rapidly converging on real-time payments, has emphasized the need for more sophisticated automation, with accurate real-time processing and intraday liquidity management. Traditional automation, integration patterns, and APIs are well understood and should be employed in various areas to derive benefits. Notwithstanding this, more can be done to increase STP rates and evolve towards intelligent automation.

AI and ML – a new reality for financial services

The remarkable rise of applied AI over the past decade has enabled computer systems to mimic human-like performance. In particular, the subfield of machine learning has the potential to add ‘intelligence’ to automation and further streamline FX operations, and such methods are increasingly becoming a reality for market participants.

While AI is a broad discipline of study, the focus is on ‘weak AI’ – or applied AI – defined as the use of computer intelligence within a defined task. ML is best thought of as a specific set of tools within weak AI, which entails the use of computer algorithms to automatically optimize mathematical models from data without explicit programming. The optimized model can then generalize and provide predictions against new data.

ML encompasses other cognitive technologies too, such as computer vision and natural language processing. Deep learning, a further subset within ML, structures mathematical models across multiple layers of nodes, mapping inputs to outputs in various combinations. This architecture is thought to mimic the structure of neurons of the human brain, giving rise to the term ‘neural networks’.

Such techniques, accelerated by the explosion of data, computer capabilities, and algorithmic innovations, are very successfully being adopted by ‘big-tech’. Furthermore, a wide range of open-source packages and cloud services are now readily available, democratizing the use of these ML solutions

Adding intelligence to STP

ML can be applied to the transaction processing lifecycle, to add levels of automation and efficiency that cannot otherwise be attained through conventional methods alone. In addition, ML offers the ability to extract meaningful insights and provide recommendations. Specific examples of ML applicability are discussed below.

Payment fraud prevention

ML can augment traditional rules-based processes in preventing fraudulent activity and/or allow for the detection of anomalous activity. When clearly categorized and labeled historical fraudulent and nonfraudulent data is available, a binary classification model can be trained and used to identify fraudulent activity in real-time when presented with new transaction data. As access to labeled data may not always be available, more sophisticated unsupervised learning techniques based on probability densities are available to detect anomalous behavior. Historical payment data can be used to ‘learn’ normal behavior, which can then be used to detect deviations from this learned normal behavior.

Nostro forecasting

Advancements in domestic and cross-border payments have necessitated proactive intraday liquidity management and funding. As a result, there is increased emphasis on real-time cash flow forecasting. ML-based cash flow forecasts can supplement conventional forecasts. Cash flows can be treated as time series data and ML methods can be applied to predict future values of the time series. Classical techniques such as ARIMA are well established, and more recently deep neural network architectures for time series analysis and predictions have also gained traction.

Predictive payment analytics

The use of predictive analytics in the payments space has the potential to offer a streamlined experience, while significantly improving customer experience. Payment failures remain commonplace and often result in significant costs. Primary points of frictions can be alleviated by applying payment data capture validation to ensure all country-specific rules are met and to validate beneficiary account details. In this regard, interoperability initiatives around ISO 20022 and payment data validation under SWIFT gpi are commendable. ML can further supplement conventional solutions and offer predictive capabilities around settlement failures prior to payments being sent, and such insights can be used to trigger additional manual scrutiny.

In relation to payment routing options, expected time to settle, and anticipated total cost, recommendations can be provided to customers that not only factor in a customer’s stated preferences, but also consider historical transactions and past selections, to suggest optimized payment routing options. These models can also be collaborative, to allow for collective input from various customers, to better drive recommendations for the benefit of all.

Confirmation matching

The affirmation and confirmation processes remain significantly human-intensive areas of back-office processing. While rules-based confirmation matching applications are widely used, paper-based confirmations and the matching thereof present time-consuming manual operations. Certainly, techniques such as QR codes on outbound confirmations and using these to look up transactions have benefits here. Computer vision can also be applied to extract data from incoming PDF confirmations and used in an automated matching process.

Rules-based systems often require human intervention to manually deal with partial matches and establish additional rules on an ongoing basis. ML can automate the suggestion of additional rules. More advanced techniques include entity resolution (or record linking) algorithms to derive matches, as an alternative to fuzzy matching logic. Finally, unsupervised learning techniques applied to incoming confirmation data have the potential to uncover previously unknown patterns and inform additional matching rules.

Reconciliation

Reconciliations, such as nostro reconciliations, are heavily rules-based. Nostro reconciliation deals with the reuniting of actual cash movements as depicted on an external statement with the recorded projected cash movements. Timely reconciliation and the use of automated processes are essential for optimizing liquidity and informing funding. As such, there is growing interest in how ML methods can supplement rules-based applications. Here too, techniques such as entity resolution algorithms and uncovering previously unknown patterns in the data can be beneficial in further augmenting traditional methods.

Harnessing data

Obtaining business-critical insights and value from data requires ready access to curated, processed, trusted, and quality data for ML and AI solutions to work effectively. Enterprises have traditionally relied on batch processing with analytics segregated from transactional applications, as is often the case with enterprise data warehouses. This is rapidly changing, with greater emphasis on streaming technologies and ML-driven predictions wherein outcomes can readily be visualized or utilized directly within the context of the transactional applications and in real-time.

Fragmentation of data across various solutions and business areas is one of the biggest challenges towards the availability of clean, consistent, and high-quality data. Therefore, investing in flexible yet secure data and interoperability architecture with streaming capabilities is key. The data architecture must support the consolidation of data and adopt best practices around master data management, metadata management, model management, data discovery, and data lineage. As ML requires large volumes of high-quality data for training to generate an optimized model through an iterative process of scoring and evaluation, data pipelines must be consistent across the complete ML development and deployment paths. This ensures equivalence between production data features and those used to train the models. ML models also require periodic retraining to make the best use of new data and to prevent model decay over time.

The recipe for success: adapt and adopt

As a specialized discipline that brings together computer science, mathematics, and statistics, ML requires collaboration across data scientists, data engineers, domain experts, and conventional development teams. A robust data and technology architecture strategy is needed to derive maximum benefits.

The explosion of cloud computing in recent years has allowed for many technology paradigms available as a service. These advances have greatly democratized technology and have the added benefits of reducing capital expenditure and lowering the total cost of ownership. Applied AI has come of age and has practical benefits that can unlock business value by lowering operational costs and driving additional revenue growth. An effective, proactive approach is best considered as part of an over-arching business transformative strategy with an emphasis on solving real-world business problems and delivering value through constant customer engagement.

While some areas of automation boost STP, ML has the potential to transform operations more broadly and drive intelligent automation. When democratized, the maturity and sophistication of ML have the capability of furthering all industry players, regardless of size, location, and current industry standing. Additional areas, which have been highly successful within the big-tech platforms and which can be applied to FX, include personalization and recommendations, conversational AI in communications, customer churn predictions, and document processing.

In today’s digital environment, continuous innovation must be embedded into the fabric of how an organization operates. The core systems and data architecture must be flexible to accommodate a future that continues to evolve at pace. The resulting business agility allows for tackling the constant and competing pressures from customers, regulators, and shareholders alike.

This article was originally published in e-Forex