This is the continuation of the multi-part blog series on Emergent tech that are poised to drive the financial and banking domain in the immediate future. The fifth part of this blog took a look at advantages of RPA within the finance, the difference between AI-driven RPA and conventional RPA, the key capacities for cognitive automation, and the part of low-code development. 

The sixth part of this multi-part series of this blog series will look at further use cases for AI within the finance and banking industry. 

Client-facing chatbots 

Chatbots are considered to be VAs (Virtual Assistants) that assist client transact or identify resolutions to problems. These automated programmes leverage NLP (Natural Language Processing) to interact with customers in natural language (through voice or text), and leveraging ML algorithms to get better with the passage of time. Chatbots are being put out by an array of financial services firms, usually in their mobile applications and on social media. While several are still in the trial stages of development, there is prospective for growth as chatbots witness increased proliferation and utilization, particularly within the younger generations, and become more advanced.  

The present generations of VAs and chatbots that are in deployment by financial services enterprises is simple, typically furnishing balance data or alerts to clients or responding to basic questions. It is worth observing that the increased proliferation and usage of chatbots is connected with the increased utilization of messaging applications.  

Chatbots, on an increasing basis, are shifting in the direction of providing advice and prompting clients to act. On top of helping clients of financial enterprises in undertaking financial decisions, financial enterprises can gain advantages by obtaining data with regards to their clients on the basis of their interactions with chatbots. While legacy infrastructure for client data storage has reduced the speed of development of chatbots within financial institutions in several jurisdictions, Asian financial institutions and regulators have produced more advanced chatbots that are presently actively being used. The insurance space has also looked into the utilization of chatbots to furnish real-time insurance advice. 

Operations-oriented use cases 

Financial enterprises can leverage machine learning and AI technologies for an array of operational (or back-office) applications. A few of these applications consist of: (i) capital optimization by banks, ii) model risk management (back-testing and model validation); and iii) market impact analysis (modelling of trading from big positions.) This portion of the blog takes a look at each of these in turn. 

• Capital optimization use case 

Capital optimization or the maximisation of the bottom line with limited capital, is a conventional function in operating a bank that is mostly dependent on mathematical strategies. AI/ML tools developed on the foundations of computational capacities, big data, and mathematical ideas of optimisation to improve the precision, efficiency, and quickness of capital optimization. Optimization of a banking firm’s regulatory capital with ML has been a subject of interest from both a business and academic perspective, and has been so over the course of the previous few years.  

In 2012, a specialist from the private sector observed that a majority of banks stated that they had carried out meaningful programmes to optimize risk-weighted assets (RWA) and had observed 5 to 15% RWA savings. Capital optimization is additionally being performed in the sphere of derivatives margin optimization like margin valuation adjustment (MVA). New regulations with regards to clearance and bilateral margining have escalated the demand for advanced strategies for optimization of capital and initial margin. 

AI/ML could help banking enterprises in the optimization of MVA, and latest research indicates that work is being carried out in this sphere. In the context of MVA optimization, ML attempts to minimize the preliminary margin for derivatives by a combo of: a) executing pairs of offsetting derivative trades; b) carrying out offsetting techniques with the same dealer; c) Innovating trades from a dealer’s portfolio to another. 

ML witnesses the ideal combo of the preliminary margin reducing trades at a provided time on the basis of the degree of initial margin reduction, historically, under differing combos of those trades. A probable implication of these advances in MVA and RWA is a reduction in the conventionally regulatory capital and increased reliance on the non-optimisable capital regulatory utilities. 

• Model risk management (back-testing and model validation) and stress testing 

Practitioners and academics usually cite model validation and back-testing as spheres where advancements with AI/ML will probably be usually visible. Banks are looking at ML to furnish meaning to massive, unstructured, and semi-structured datasets and to handle the output generated by primary models. Back-testing is critical as it is conventionally leveraged to assess how well banks risk models are performing. In recent years, European and American prudential regulators concentrated on back-testing and validation harnessed by banking enterprises by furnishing guidance on model risk administration. Leveraging an array of financial setting for back-testing facilitates for consideration of shifts in market behaviour and other such trends, hopefully minimizing the prospect for underestimating the risk in these scenarios. 

A few applications are currently live. For example, one global investment/corporate bank is utilizing unsupervised learning algorithms in model validation. Its equity business has leveraged this variant of ML to identify anomalous projectors produced by its stress-evaluation models.  

A minimal fraction of these computations are radical, and knocked out of the typical distribution of outcomes by an abnormality of the computational cycle or incorrect data inputs. Unsupervised learning algorithms can help model validators on the ongoing oversight of internal and regulatory stress-evaluation frameworks, as they help in determination of if those frameworks are performing within acceptable tolerances or deviating from their intended purpose. They can additionally furnish input to operational risk frameworks: like the vulnerability of enterprises to cyber-attacks. 

Likewise, AI/ML strategies are additionally being applied to stress testing. The increased utilization of stress testing following the financial crisis has put forth hurdles for banking enterprises as they function to undertake analysis of massive amounts of data with a massive financial enterprise to produce tools to help them in modelling their capital markets business for stress testing in banking. The utilities produced intend to restrict the number of variables leveraged in scenario analysis for the loss provided default and likelihood of default models. By leveraging unsupervised learning strategies to review massive amounts of data, the utilities can record any biases related with the assortment of variables, therefore leading to improved models with better transparency. 

Market impact analysis 

AI/ML can supplement traditional market impact models. Enterprises can leverage AI/ML to gather more data from sparse historical data, or assist in identification of non-linear relationships in order flow. AI/ML can be harnessed to develop ‘trading robots’ that then instruct themselves how to act to market changes. Market impact analysis consists of assessing the impact of an enterprise’s own trading on market prices. Some enterprises are worried about the effect of trades, particularly of massive trades, on market prices, a more precise estimation of this effect is critical to restricting to the timing of trades and reduction of trade execution expenditure. 

Firms are looking into leveraging AI utilities to evaluate the market impact of a provided trade. The impact of an enterprise’s own trading on market prices is ridiculously difficult to model, particularly for far less liquid securities, where data on comparable historical trades are scarce. AI utilities might assist through augmentation models currently in utilization, or by putting forth a ML strategy to reduce trading effect on prices and liquidity. For a majority of the most active systematic funds, as much as 2/3rds of the gain on trades are predicted to be lost to market impact expenditure. AI utilities might assist through augmenting models already in utilization, or by putting forth an ML strategy to reduce trading impact on pricing and liquidity for trading both into and out of large market positions, or as a part of daily trading techniques. 

ML is usually leveraged to detect groups of bonds that act in a similar fashion to one another. By doing so, they can be reliant on several more data points, furnishing improved estimates of pricing movements when the market is thin. The resulting utility groups bonds into broad, intuitively similar buckets and then, leveraging cluster analysis, gathers the most comparable products together in every bucket, in scoring the liquidity of each individual bond. 

Additionally, AI/ML can be harnessed to assist in detection of how the timing of trades can reduce market impact. Market impact frameworks can be produced that detail how the impact of a trade is dependent on prior trades as a starting point. The models make an effort to prevent scheduling trades too closely together to prevent having a market impact bigger than the sum of its parts. These frameworks can be leveraged to establish the ideal prospective trading schedules for an array of scenarios and then tweak the schedule as the actual trade goes forward, leveraging supervised learning strategies to make the short-term predictions in determination of those tweaks.  

Banks are also evaluating reinforcement learning to teach AI/ML utilities to react to order imbalance and queue position in the limit order book.