Forward-thinking financial institutions already understand that properly deployed AI can yield worthy benefits – not just when it comes to making the sector more efficient, but in providing a more intuitive, engaging user experience. According to a Business Insider report, 80% of banks understand that AI technology is fast becoming standard-operating-procedure. A similar number are already investing in research and acquisition programs designed to adapt to the obvious future.
The centrality of AI technologies to the financial sector today is also forcing developers and business managers alike to re-think the way they refer to new developments in the industry. A decade ago, any consumer-facing application that dealt with finance was referred to as part of the “fintech landscape.” Today, the division between “fintech” and “finance” is all but invisible – DevOps approaches are now used in banking and financial institutions and examples of AI in fintech are all around us.
In this article, we’ll look at this emerging confluence and the reality that the financial and technology sectors have grown ever more indistinguishable. The case studies below represent, to our mind, some of the most exciting new developments.
Adoption and Adaptation
First, let’s take a broad overview of the ways in which AI is making an impact in the fintech space. Perhaps the most obvious of these developments is chatbots, which have so exploded in popularity over the past five years that an entire sub-industry has grown up around them. They are now a real part of the landscape of financial advice available to consumers. This industry is still growing, however, and is predicted to be worth over $17,440 million (with a CAGR of 17.9%) by 2027, by which time chatbots will have logged an estimated 862 million hours. These are hours saved from the schedules of live customer service agents.
The strange thing about the explosion of chatbots is that they were originally conceived as one part of a range of AI-driven capabilities. The programming frameworks which are used to train them, in fact, are now being used to facilitate the development of much more advanced AI capabilities. As we’ve previously reported, some of these are focused on hyper-personalization. Others work to achieve business efficiencies by improving resource management.
Others are even more advanced. Visa, for instance, uses AI tools to spot fraudulent transactions in real time and claims that this technology is able to detect up to 500 unique transaction risk attributes. According to Visa, for 2019, the share of fraudulent operations on a global level in the Visa system reached a historical minimum and is less than 0.1%, due in large part to this system.
Other financial institutions have focused on other applications of AI technologies. Bank of America, for instance, released Erica in 2016 - a digital assistant that is integrated into the bank’s mobile banking app. Though superficially resembling a chatbot, the assistant also leans on some of the deeper capabilities of the AI frameworks mentioned above. It is able to access third-party systems on a user’s behalf and to alert them to changes in their credit rating. In addition, Erica incorporates Natural Language Processing (NLP) to allow users advanced search capabilities, such as finding particular transactions and tagging payments which are more expensive than expected using ML-driven anomaly detection.
Barclays and AI Simudyne
The most advanced recent instance of financial institutions looking to leverage the power of AI, however, has come via the partnership of Barclays (a UK-based bank), and AI Simudyne, a startup focused on agent-based modeling.
Simudyne’s offer was originally focused on providing hyper-personalization to fintech startups. However, since their partnership with Barclays, they have extended the system with the aim of providing real-time intelligence to both banking analysts and customers alike. This has been so successful, in fact, that the system is now able to offer customers financial products autonomously.
Technical details of the system remain protected, but it’s clear that what AI Simudyne has managed to do is deploy the same AI tools that are used to respond to user requests to model the behavior of users in a much broader way. They have suggested that the system built for Barclays aims to simulate the relationship between a lender and a client and to assess the risk of lending to them based on this simulation. Iterating this system then allows Barclays to assess the risk represented by new products or new markets.
This move shows how existing tools can be extended to novel environments. The data collection, training, and modeling AI systems which are already used in chatbots and personalization software, it seems, can be extended to autonomously make business decisions, as long as the managerial appetite exists for this.
Though the examples of AI in fintech we’ve discussed above are certainly impressive, the details of how these systems work are generally hidden. Several companies, however, have released detailed descriptions of systems built to achieve the same kind of outcomes as those outlined above – to detect fraud or to perform automated stock portfolio management. In this section, we’ll take a look at two examples of this.
One of the most detailed explanations of the way in which AI can be used in financial fraud detection comes from Altexsoft. They first make the point that ML-based fraud detection is far superior to the systems it has replaced, which were mainly based on prescriptive rules. ML systems, trained on real-world data containing instances of fraud spotted by human analysts, have become quite successful.
Feedzai, for instance, claims that its ML system can detect up to 95 percent of all fraud and minimize the cost of manual reconciliations, which accounts now for 25 percent of fraud expenditures Capgemini similarly claims that fraud detection systems using machine learning and analytics minimize fraud investigation time by 70 percent and improve detection accuracy by 90 percent.
These systems use a variety of approaches to fraud detection:
- Anomaly detection. This is one of the common anti-fraud approaches in data science, and is based on classifying all objects in the available data into two groups: normal distribution and outliers. Outliers, in this case, are the objects (e.g. transactions) that deviate from normal ones and are considered potentially fraudulent. In fintech environments, this data may range from transaction details to images and unstructured texts.
- Supervised learning entails training an algorithm using labeled historical data. In this case, existing datasets already have target variables marked, and the goal of training is to make the system predict these variables in future data.
- Random forest trees are a more supervision-heavy form of fraud detection. In this model, data is iteratively sorted into a pre-defined number of categories. Human analysts are then able to review the transactions in these categories for potential fraud.
- For more complex datasets, there are more advanced models. A support vector machine (SVM), for instance, is a supervised machine learning model that uses a non-probabilistic binary linear classifier to group records in a dataset. It does this by cutting hyperplanes through the phase space made by the available data and then choosing the hyperplane that separates data most distinctly as the basis for its fraud isolation.
For developers, all these approaches may be of technical interest. For banks, however, the value of fraud detection comes down to economics. Just as AI can be used to secure digital assets, it’s now clear that it can also be used to identify potential write-off losses before they can cause monetary damage.
The other side of the coin in reducing the potential for financial fraud is to use AI systems to improve the performance of investment portfolios. As we’ve discussed above, Barclays has made use of a system of this type in order to assess the risk of lending to certain customers; others have made use of the same techniques to offer fully automated investment advisors designed to realize the maximum possible gains, all based on client’s preferred investing style.
And this is not just theoretical. Carbon Collective co-founder, Zach Stein, credits the “robo-advisor” concept for the startup success of his climate change focused investment management company. “Our algorithm has not only allowed for cost savings via a smaller staff but removes human error and emotion from the process. Our automated client portfolio management has allowed us to outperform the overall market in recent years.”
Some companies have been spectacularly successful in doing this. Betterment, for example, offers customers a range of services – they can either let the robo-advisor execute trades on their behalf or merely use this as an advisor to guide their own investing decisions. At a broader scale, we have recently written about using Pony in a fintech environment to build high-performance tools. Pony is a new actor-model language that's statically typed and ahead-of-time compiled (using LLVM), with a fully concurrent garbage collector and a data-race free type system, and is seeing significant use as a platform for building investment management AIs.
Again, the details of how robo-advisors work is proprietary and closely guarded. However, it’s possible to speculate about the components that are required to build such a system:
- First, the client is asked to provide personal financial data – either by filling out an online form, or adding existing accounts to the system through an appropriate API.
- The system will then add the data to its Data Lake using secure distributed storage. HDFS (Hadoop Distributed File System) is the standard of choice in this instance, as it allows data to be received and processed from various systems with different formats and infrastructures.
- The software then uses specific algorithms to form, manage, and optimize the client’s portfolio. As the conditions of financial markets change, the portfolio is automatically rebalanced to reflect this. When necessary, the system performs tax-loss harvesting operations to minimize tax liabilities.
- At this stage, it’s possible to deploy a broad range of AI systems in order to optimize the decisions being made. In fact, one of the advantages of this kind of system is that it allows any AI engine to be plugged in and used to manage the data generated by the steps above.
- Upon processing the transactional data, the software feeds it to the Enterprise Data Warehouse (EDW), a centralized database available for analytics and broader use.
- The client’s investment history, portfolio progress, risks, and other relevant information can be visualized with the platform’s dashboard.
As such, this kind of system is based upon a number of nested sub-components, many of which are available as stand-alone systems. Though developers will need to pay careful attention to the way in which these components interact, it is technically possible to build an AI-driven investment advisor using pre-built components.
This is not to say, of course, that building robo-advisors of this type is without risk. As with any technology in the fintech space, the security of these systems should be carefully designed and continually assessed, and developers should take care that their eagerness to deploy ML models does not overpower their concern to keep data safe and secure.
These are, of course, just two of the ways in which AI may be used in fintech. There are plenty of other exciting developments in the space. Some developers are now using chaos testing to improve the stability of fintech products and there are always tinkering about how to best incorporate blockchain.
In other words, AI is just one of a number of emerging technologies that are going to transform the way that the financial industry operates over the next decade. And, as these technologies become increasingly integrated into the industry, we will see the divide between tech and finance become more blurred than ever before.