As the cost of doing business rises, so does the need for investment banks to cut costs and increase revenue while providing the highest level of service to their clients.
Machine learning and artificial intelligence (AI) represent a huge opportunity for investment banks to improve their highly diverse set of businesses – but how can they get started?
A new eBook highlights practical use cases for AI in today’s investment banking market.
Armed with this knowledge, investment bankers can take advantage of the enormous amount of data they generate and transform into AI-enabled enterprises.
The level of interest in automation, optimization, and artificial intelligence (AI) in investment banking is unprecedented. Driven by the increasing cost of doing business, the accelerating innovation of startups and the ever-decreasing margins created by technological advances, the level of interest in AI is only exceeded by the level of confusion about how best to take advantage of the technology.
YOU MAY ALSO READ THIS: 7 Ways Sterling’s Blockchain Commodity Marketplace will Disrupt Agro-Commodity Supply Chain
Cutting-edge machine learning technology plus the large troves of data generated every day by investment banks is a potent combination, but how best to harness this power?
A typical investment bank contains a highly diverse set of businesses providing advisory, transactional, research, trading and other services to corporate clients, institutional and individual investors, and other market counterparties.
Investment banks can use AI in six critical ways:
RFQ pricing optimization: In markets where requests-for-quotation (RFQs) are the usual way of transacting, there is often a trade-off between maximizing the likelihood of printing a trade and maximizing the bank’s margin. Should traders go for volume, keeping spreads tight and winning a high share of RFQs sent their way?
Or should they be more selective, aiming for a lower number of more lucrative trades? Are traders leaving money on the table by offering tighter spreads to counterparties when these are not needed to win the trade? Using AI, trading desks can model the likelihood of winning any individual RFQ at any given spread over a reference price, considering factors such as the nature of the counterparty, their historical trading behavior, which security is being traded and at which size, even concurrent market action.
This allows trading desks to quickly determine how aggressively to price their quote by rapidly predicting how likely the requestor is to trade at on a curve of possible bid/offer spreads, and to dial in the optimal spread in terms of margin for the bank
Algorithmic trade execution: . As the number of trading venues proliferates and the frequent domination of liquidity by high frequency traders continues, the requirement to demonstrate best execution to relevant regulators has become crucial. Sophisticated algorithmic trade execution strategies are now an important component of execution and flow trading businesses, even for those high-touch trades which are negotiated by telephone.
These algorithms determine how best to route or execute a trade to get the best possible result for the firm or for the client. Predicting the best order routing and execution strategy frees traders up so that their time can be spent finding profitable opportunities for client flow and revenue generation. Based on historical algorithmic trading transaction cost reporting and analysis (TCA) data, banks can train an algorithm to predict the best order routing to optimize price and cost and minimize market impact.
Non-linear machine learning algorithms can add a sophisticated extra layer of predictive analytics to traditional rulesbased smart order routers (SOR). A machine learning model can predict the best routing, pace, trading venue, counterparty, and legal entity for each individual trade to get the best possible price at the lowest possible cost, even picking from multiple available rules-based models to choose the most appropriate one for a particular trade at any given time.
Machine learning techniques can also recommend the best execution strategy. Based on prior trades and market data, machine learning algorithms can pick from, for instance, with-volume execution, order slicing, VWAP targeting or a mixture of these, even predicting the shape of the volume-over-time curve to minimize market impact. In addition, real-time AIdriven recommendation of individual order size, pace and price limit, and price trend data is also possible, with the trading algorithm becoming more aggressive in favorable market microconditions and less aggressive when the market is becoming unfavorable (or, indeed, vice versa, depending on the trader’s goals and views).
Research recommendations: In today’s post-MiFID2, commission-unbundled world, banks are having a long, hard look at how they provide research to clients. The days of “paid per page” seem like a distant memory. Whilst providing equity, credit, economic and asset allocation research continues to be a mainstay of investment banks’ client-facing business, the trend towards enforcing the separation of research payments from trading commissions means that ensuring that the right research notes get put in front of the right clients at the right time is now more important than ever. At the same time, the typical institutional investor is faced with a barrage of sometimes hundreds of research reports a day, many of which end up unnoticed and unread.
Reducing the noise by sending only timely and relevant research adds credibility to the bank’s brand. Ultimately, banks will probably end up sending out far fewer research reports, but readership rates (and hence revenues) will be dramatically increased.
In addition, the process of building such models identifies the key drivers of research consumption, and these insights can be used to increase overall quality and relevance of research output. Fortunately, the data for building such a research recommendation model is often at hand.
Research departments can build more relevant content, and sales desks can ensure that the product gets highlighted to the right investors at the right times. Trade and quote history, industry, employer, timing, newsflow timing and many other attributes can all be considered as inputs to identify the right recipients
Operational break and failure prediction: Investment banks’ operations departments contain many complex processes whose steps and outcomes are painstakingly and completely recorded, with large quantities of such data generated every trading day. Over the last 25 years, automated techniques such as straightthrough processing (STP) have substantially reduced operational costs.
But as operational processes have become more complex and cycles become ever faster, errors have become most costly to find and fix before they propagate, or worse, the client becomes aware of them. By some estimates half of middle and back office cost is attributable to detecting, researching, and correcting operational errors and exceptions.
Banks can improve service quality and speed, while reducing operational costs, if they can predict and avoid, or detect and fix, operational errors and exceptions before they become costly. Machine learning can be used to predict when operational errors will occur, due to exceptions requiring special handling, unsettled trades, breached collateral or margin limits, broken reconciliations, or pricing errors. Knowing where errors are likely to occur allows middle and back office staff to act preemptively to assure operational problems are avoided.
Of course, not everything can be reliably predicted. Machine learning can also be used to detect operational errors that have already occurred. Using both supervised and unsupervised learning in tandem can be a powerful way to detect such errors. Supervised learning can be used to train a model to detect errors based on errors that have occurred in the past.
Multiple models can be trained with automated machine learning to spot different types of errors with high confidence. Unsupervised learning techniques such as anomaly detection can be used to identify situations that may indicate a problem – – such as a price that looks unusual.
Trade and communications surveillance: Rogue traders represent a serious and difficult to identify risk. The most infamous rogue traders have wracked up multi-billion dollar losses before being caught, terminated, and in some cases convicted and jailed. To combat such market abuse, manipulation, and fraud, as well as for risk management reasons, regulators require that investment banks monitor their traders and other employees’ trades, behaviors and communications.
Typical rule-based monitoring mechanisms detect trades with unauthorized counterparties, to unauthorized accounts, those without established trading limits, or volumes or exposures above an agreed limit. But more sophisticated mechanisms are needed.
Banks can use machine learning to identify trading patterns that are unusual for a trader based on their historical trades, or that are unlike similar traders’ activity patterns. These anomaly detection algorithms can leverage a banks own data to spot patterns of activity that may require greater scrutiny.
AI can improve both the efficiency and the quality of this surveillance while also reducing the cost of false positives by using outcome data from past investigations to flag which investigations are more likely to find actual suspicious activity. Modern outcomebased text analysis techniques can be used to incorporate electronic messaging and even voice communications to increase accuracy even further — thus making AI a powerful magnet to help find the needle in this particular haystack.
Regulatory due diligence: Every year banks spend millions of dollars on onboarding and surveillance processes to ensure compliance with anti-money laundering legislation – and for good reason. It’s not uncommon for regulators to levy fines for inadequate or lax anti-money laundering (AML) monitoring in the order of hundreds of millions of dollars or euros.
Investment banks depend on their compliance teams to ensure that regulatory and reputational risks are avoided by stringently following anti-money laundering regulations, including exacting onboarding processes and regular refreshes of documentation checks. Machine learning can help banks determine which aspects of their know your customer (KYC) program actually correlate with potential money laundering risks, enabling them to reduce friction in the new client onboarding process, and which data points may provide triggers for enhanced due diligence (EDD) investigations.
Such techniques not only increase a bank’s ability to target the right clients for heightened scrutiny but also supply a quantitative justification for regulators, in sufficient detail that reasons for individual investigation triggers can be provided for each EDD.
These can also be used to proactively recommend extra handholding for such clients to ensure a smoother process.