Published: August 24, 2021
Machine learning is often seen as the panacea for all modern business challenges. However, while it can be a very useful ally, it won’t fix everything – treasurers will still need to make judgment calls. In the company of seasoned experts, TMI explores the reality underpinning this much-discussed technology.
The current level of conversation around machine learning (ML) might suggest that it has already become one of the most prevalent treasury solutions. The truth is that while it is making significant headway in terms of adoption in the treasury context, the journey has only just begun; many firms are still working out where it might fit, and how it can be best used. Is it really time for treasurers to engage?
“Over the next three years, we will create more new data than we have in the entirety of human history combined,” states John Pizzi, Senior Director, Enterprise Strategy, Capital Markets, FIS. He argues that artificial intelligence (AI) is “no longer a science fiction”, citing PwC figures that show AI is expected to contribute nearly $16tr. to the global economy over the next decade.
Few can fail to have noticed the use of AI by Big Tech firms such as Amazon and Google in helping or steering our decision-making. But in both organisations, the use of AI goes beyond what is obvious to consumers. Amazon is using it to optimise its logistics and warehousing operations, and Google is developing AI to aid screening evaluations for lung cancer patients.
The fact is that the impact of AI is felt across a far wider scope of industries than many would imagine, and in the financial services sector it is already strongly evident. Pizzi notes that around 44% of capital markets firms are already using AI in their trading processes, a figure that he says will rise to 61% by 2023.
“Machine learning is being used to make prediction models more accurate, and in quantifying and reducing credit risk,” he continues. “Predictive analytics are producing more accurate insights around cash management and fraud, and automation is transforming the workplace, improving the speed, accuracy and efficiency of operations.”
In practice, ML is “a piece of software capable of flagging data to match specific patterns”, explains Pedro Porfirio, Global Head of Capital Markets, Finastra. In some cases, instead of using pre-defined rules, it can learn by itself what these patterns are. It needs large volumes of data to be effective, and it can process information and detect patterns much quicker than a human being.
Pedro Porfirio
Global Head of Capital Markets, Finastra
“Treasurers need to navigate a lot of data to make decisions and these decisions are made given the patterns of the data, so a good use case is around fraud detection,” explains Pizzi. “As ML use starts to become more widespread, treasurers will find it useful to analyse their cash flow data, and to support cash forecasting and liquidity management. While ML technology continues to evolve, those who are quick to adapt will find themselves ahead of their competitors.”
It’s a compelling scenario, but for many treasurers the business case for adoption is yet to be heard. To tackle this adoption latency (and benefit from Pizzi’s predicted competitive advantage), at the very least the business case must reflect the problem being solved and objective being met, says Bruce Meuli, Treasury Advisor Executive, Bank of America (BofA).
Most business cases tend to start off with a cost-efficiency or reduction, he notes. The business case around some early applications of ML has therefore been around reconciliation processes. Here, ML, as a set of “smart algorithms”, is taught to understand related values of an input (or multiple inputs) and then make a best-match based on that understanding.
Cash application is a classic use case in this scenario, says Meuli, explaining that BofA has developed a solution for analysing historical data and events, identifying patterns, and taking a “largely supervised” (requiring human teaching) learning approach to defining output.
With very little ‘unsupervised’ (or self-learning) ML currently in treasury, he explains that human insight is used to ‘train’ the ML models to follow set business rules. Output will thus be a response that either agrees with the rules and which processes accordingly, or detects an exception for further human action.
Recourse to human intervention may appear to defeat the object of ML but, says Meuli, when it comes to the business case, this step can be seen as part of the preliminary set-up. A first-pass match within cash application, for example, may consider customer, invoice number, amount and date, and initially achieve a 60% matching success. The next stage of set-up may enable use of additional information (such as remittance data and customer grouping) with human intervention to refine that understanding.
So while the initial business case focus is on cost reduction and decision support, a phased approach, even with some manual intervention, enables the system to scale up over time, progressively moving to more complex data flows and eventually non-human judgments.
Kunal Makin
Director, Global Treasury Data Product, Bank of America
The adoption of ML does not have to mean the removal of human input. Far from it, argues, Kunal Makin, Director, Global Treasury Data Product, BofA, the human/ML interface can deliver some impressive results that can lead to better advisory from banking partners and richer insights.
ML has gone through a complete evolution cycle since the early days of elementary analysis and modelling in the industry, says Makin. However, while as an automated means of identifying and leveraging patterns in data it has real power, execution of the cognitive work, such as pre-processing the data and choosing the right parameters, must be carried out by humans.
“While human intervention has exponentially decreased in the models that have more predictor functions, ML-like analyses which provide strategic outputs to drive both top and bottom line value in a business are very much human dependent even today,” he notes. “This remains true across treasury functions where ML is used, such as in liquidity forecasting, AML and fraud management.”
Of course, ML outputs can vary depending on the business inputs which are human dependent. For Makin there’s a whole host of “human-in-the-loop” ML approaches, where humans are called to make a final decision when the algorithm is uncertain and the result of the human decision is then used to improve the model. This, he explains, is why output variables are also very much dependent on human interpretation and usage of the model.
It is critical that the business problem is well-defined at the onset of model building so output parameters can be appropriately plugged-in to solve the problem, comments Makin. “It’s also important to consider that while ML accounts for the ‘standard’ data points within the 80% of the bell-curve of a typical problem, there will always be those outliers making up the 20% of use cases at the bottom and top where individual human assessment of the applicability of model to the business scenario is critical,” he notes. “Given the recent changes in the macro environment driven by the pandemic, we encounter these outliers quite often.”
The beneficial application of human intelligence in the assessment of outliers suggests that hybrid models may be applied to most treasury use cases, continues Makin. This includes payment optimisation, liquidity forecasting, next-product lead generation, FX analytics, credit risk and underwriting for supply chain activities.
That said, he acknowledges that there are certain functions in the treasury context, such as bank onboarding, that are currently best served by ML-based robotic process automation (RPA) tools. “These are processes that are repeatable and that would otherwise generate a margin of error with human intervention,” he explains. “We continue to work on identifying process improvement opportunities across treasury which have a very low human touchpoint and which create efficiencies for both our clients’ experience and us.”
Kunal Makin, Director, Global Treasury Data Product, BofA
There are different pathways open for ML adoption within treasury. Most users will take a sub-process, such as matching, which can be deployed through a third-party vendor system. “There is no need to bring every system up to the required level in one move,” says Meuli. “Treasury can target a single unsatisfactory process and fix that before moving on.”
However, there is one stage that cannot be ignored, he adds. The most important aspect of every ML project is data quality. “Only when the data is clean will it be possible to think about the tasks it will be used for, the kind of modelling and analysis that will be needed, and the systems and expertise that will be needed to make best use of it.”
Most treasurers will recognise that data de-duping and cleansing projects are rarely quick, and thus may constitute the largest proportion of project time. In the context of ML and its aims, such a project also requires a deep understanding of the data required, and a clear-sighted view of the end goal. It will be useful to talk to business colleagues to help define that goal so that the data can be formatted in a way that will be both easiest to use and most beneficial to the company.
While the promise of ML may generate excitement, it’s important to stay grounded. The target problem may be resolved with a relatively simple automated input. When considering the right tools, the caveat of avoiding the use of technology for the sake of it applies. A well-defined set of aims for ML will help define the shopping list. Indeed, advises Meuli, “starting with the problem, not the solution, is the rational way forward”.
Be warned though, he continues, “an unquestioned belief that technology will do it all is often a fallacy”. At the very least, he says every system needs to be set up and managed, and to this end some of the best solutions are where technology integrates with humans who will manage the exceptions and the teaching and learning process.
Elon Musk may believe that “robots will do everything better than us”, but is this correct? While humans cannot match ML capacity to endlessly crunch data and detect patterns, the level of knowledge, experience and intuition that treasurers bring to bear on many more nuanced situations cannot be replicated with full confidence, yet.
FIS’ Pizzi admits: “People are still very much at the centre of the workforce.” But he acknowledges that “over time, human jobs are going to change”. For him, the future is encapsulated by Amazon Chief Technologist, Tye Brady, who has said it will be “a symphony of humans and machines working together”.
In practical terms, an ML model that is proven to give high-quality results and drastically improve workflow efficiency is something that treasurers should think about. But we’re a long way from AI reaching the level of human intelligence and replacing the entire job of the treasurer.
That said, the capabilities of ML will continue to improve. The focus today for treasurers should be on understanding how humans can complement such improvements, and using time released by automation to add value.
Of course, there is a matter of confidence in technology to take on vital roles. Computers can fly and land a commercial aeroplane perfectly well, yet few passengers would feel happy without a well-trained human pilot on board. Similarly, where the finances of a company can be a matter of existential importance, few boards would hand over full responsibility to a machine, expecting human expertise to correct errors or take over when needed.
As such, Meuli feels it will be the non-critical, mundane and repetitious operational jobs that will be the first to go. Typically, this means low-complexity back-office tasks with business-rules-driven workflows. Once a business has satisfied itself that this is a feasible approach, he suggests that more advanced decision-support tools, such as the aforementioned cash forecasting solution, may be given greater authority, even if the output at this level will still be human-reviewed.
Porfirio believes that one of the most important skills a treasurer offers is the ability to make nuanced judgment calls. “AI and ML will never be able to fully replace that process. What it does offer is the ability to analyse vast amounts of data in a way that a human would not be able to, revealing patterns of interest. It is then up to the treasurer to evaluate and determine the next steps.”
Front-office tasks can be managed in this way too, but Meuli says these roles can still require considerable human involvement. That said, the banking community has invested vast sums in enabling technology to take on some critical processes, but he agrees that these have taken a long time to develop.
Indeed, computational finance for stock and bond pricing dates back to the 1930s, and in the 1950s it was first used for portfolio selection. Today, while algo trading systems calculate and transact billions of dollars of trade automatically in fractions of a second, the solutions capable of doing this are proprietary tools developed by and for specialist financial institutions.
“Typically, adoption of ML is not about completely replacing but taking away certain elements of a process,” explains Meuli. While it may do all of the initial work, faster and more accurately, even making some basic decisions, active decision-making based on prescribed output will, he contends, remain a human task for now.
Bruce Meuli
Treasury Advisor Executive, Bank of America
In the short- to medium-term, while service centres for accounts payable/accounts receivable (AP/AR) may see a significant reduction in employee numbers at the hand of automation and ML, most professional treasuries are already limited in number and arguably will not be displaced by machines. Meuli is adamant that there are “far more positive aspects to having ML working within treasury than there are negatives”.
A common argument for the control of AI-based tools has its roots in legal and ethical discussion. FICO, a global analytics software firm, recently released its State of Responsible AI report. Working with market intelligence firm Corinium, it found that despite the increased demand and use of AI tools, almost two-thirds (65%) of respondents’ companies can’t explain how specific AI model decisions or predictions are made.
The study also found that 39% of board members and 33% of executive teams have “an incomplete understanding of AI ethics”. While compliance staff (80%) and IT and data analytics teams (70%) have the highest awareness of AI ethics and responsible AI within organisations, wider understanding of these points remains minimal.
“Over the past 15 months, more and more businesses have been investing in AI tools, but have not elevated the importance of AI governance and responsible AI to the boardroom level,” comments Scott Zoldi, Chief Analytics Officer, FICO. “Organisations are increasingly leveraging AI to automate key processes that, in some cases, are making life-altering decisions for their customers and stakeholders,” he says. Senior leadership and boards “must understand and enforce auditable, immutable AI model governance and product model monitoring to ensure that the decisions are accountable, fair, transparent, and responsible”.
So is this a warning shot across treasury bows? Meuli thinks not. “Treasury is not dealing with decisions that impact people directly, but there may be considerations around how ML is used to treat suppliers, for example.” Indeed, he adds, the switch to ML simply replaces an individual’s actions with that of a machine, and individuals harbour bias, so wherever there are ethical issues in current treasury practice, the same will apply when ML is deployed.
That said, he shows some concern for potential reputational damage arising from misuse of ML, mainly as a result of acting on inaccurate output. Banks such as BofA employ large teams just to audit models and ensure data integrity, but at the treasury end of the spectrum the need to understand the often inscrutable ‘black box’ nature of AI/ML is vital.
Referring to the Nobel Prize-winning Black-Scholes mathematical model for pricing options contracts, Meuli believes that few who use it really understand how it works beyond its common input variables. “Using ML models that are making decisions, yet not really understanding how they do so, or failing to implement proper auditable processes around how those decisions are arrived at, could expose an organisation to reputational damage,” he warns. With the legal and ethical debate around AI/ML in its infancy businesses should expect more conversation, and maybe even auditing, in the future.
“There is a correct way of approaching ML,” counsels Meuli. “The best results come from combining the right tools with a mindset that starts with a specific problem and objective before looking at how it can assist.” This is where the currently marginalised discipline of Six Sigma process review and improvement can be advantageous, he says, noting that it is making “something of a comeback”.
“It’s about evaluating your processes and revealing which are high volume, where data is most intensively used, where and how decisions are made off the back of this, what the risks involved are – including reputational risk – where manual intervention is required, and then judging whether some or all of this could lend itself to ML.” The most successful corporates in this space, he notes, will use more than one tool to achieve their objectives, “and that includes humans”.
As might be expected, external resources would include vendors in this space. “ML can bring immense benefits to treasurers, but to take full advantage of it some changes might need to happen,” says Finastra’s Porfirio.
“Where the development and deployment are done on site, it requires a high level of technical skill and commitment of resources,” he says. “It will require setting up the environments, accessing data –probably including publicly available information – understanding and developing ML algorithms, which can be heavy in statistical and mathematical concepts, and defining their use case.” This may seem like a major undertaking, and for Porfirio, while larger organisations might be able to afford these investments, he feels many are unlikely to make the business case.
“Moving their technology stack from a conventional on-site set-up to a modern cloud deployment will give treasurers the opportunity to accelerate innovation and time to market, and the ability to transform their operations at scale,” he suggests. Finastra’s FusionFabric.cloud, for example, enables its clients to connect to fintechs that are developing ML algorithms for several use cases, including fraud detection, price movements, and stress scenarios.
Banking partners are a valuable resource too, not least because players such as BofA have walked the ML pathway for some time, and have the expertise and client contact to help treasurers explore and understand where the technology can be best deployed. In doing so, they can help eliminate any ‘black box’-type fears that could restrict confident ML uptake.
Indeed, for Meuli, while “the end-goal for us as a bank is to help treasurers gain more business insight and work more efficiently”, it is incumbent upon it to help treasurers gain “at least a high-level understanding of what ML is capable of doing and what it is not capable of doing”. Then, by better understanding how specific ML models work, it becomes possible, for example, for treasurers to begin collecting additional data points and refining their own models over time, making the system invaluable (and validating the business case).
Through contact with ML-engaged corporate clients, banks such as BofA are also well placed to provide focused solutions. The bank is, for example, developing an ML programme based around natural language processing (NLP). This is aimed at extracting common questions and their related responses from a vast dataset of historical client documents so that the bank can more rapidly answer client enquiries.
A specific challenge in treasury is accurate cashflow forecasting. BofA has been working with a fintech to help forecast cash volumes for its clients using its CashPro Forecasting platform. This can predict a specific account balance at any determined point. It uses historical balance data analysed across a specified time series, but also takes into account variables such as seasonality or even known events, such as the pandemic, to generate realistic context-based output.
BofA is also using Big Data analytics and ML-based pattern-recognition models to enhance fraud detection, and is now offering intelligent payments routing as clients seek the most appropriate (and cost-effective if desired) cross-border pathway for their transactions.
Furthermore, bank reconciliation and cash application processing, and payment pre-validation tools are also now part of the bank’s ML-driven toolkit. Indeed, with its STP rules engine and file validation pilot, treasury clients will be able to detect issues with payments before they reach the bank. As Meuli comments, “post-execution reporting and checking should now be redundant”.
The bottom line with any technological solution is whether or not it solves the target problem. The engagement of banks such as BofA with corporate treasury clients is essential to avoid creating solutions that are merely looking for a problem.
It is this collaborative approach that seems to be fundamental to the furtherance of ML adoption and progress. It certainly goes a long way towards realising that ‘futuristic symphony of humans and machines working together’.
To find out more about this topic, see the accompanying infographic:
In many ways, Covid-19 has been the ultimate test for treasury teams and their technology. With the spotlight firmly on treasury management systems (TMSs), treasurers are asking whether they have the best tech for the job – in a remote-working environment where instant access to ‘decision-ready’ data is critical. Is there a better TMS out […]
