Applying op risk and probabilistic networks to corporate actions processing
CORPORATE ACTIONS PROCESSING
Bayesian belief networks may provide a solution for arriving at loss estimates for Basel II. By Shilpa Ramamurthy, Harpreet Arora and Anirbid Ghosh
WITH the Basel deadlines fast approaching, most affected banks are scrambling to get their operational risk management act together. Among many factors that are impeding the development of stable operational risk management infrastructures, unavailability of loss data is probably the most common. Without a credible internal loss history database, most of the advanced risk analysis and measurement techniques (for example, the loss distribution approach) cannot be implemented. The Basel-defined requirements for loss history are fairly stringent – to use any one of the actuarial approaches, organisations would need to have at least three years (preferably five) of operational loss history. Most banks and financial institutions, if not all, will not have this historical data, especially for some business lines – risk category combinations as shown in table A.
One way to get around this limitation is to use qualitative approaches such as scorecards or key risk indicators. However, these approaches depend on purely qualitative risk assessments and hence do not involve the same amount of rigour as actuarial approaches such as the loss distribution approach.
Bayesian belief networks (BBNs) provide an elegant solution to this problem. They combine both qualitative and quantitative information for arriving at loss estimates. They are particularly appropriate for modeling operational risks with little or no historical losses – most low-frequency/high-severity operational losses fall in this category. Moreover, BBN's are causal networks, unlike other approaches such as simulations, and are particularly useful for analysing causes that contribute to operational risk.
This article provides a reference implementation of a BBN with an application to corporate actions (CA) processing. It tries to model operational losses from 'missed CA announcements', a fairly prevalent problem in the custody business. Though the implementation is for a specific risk category, the technique can easily be extended to any other category.
Bayesian networks – a brief overview
A BBN, also known as qualitative or causal probabilistic network, is a technique that helps model, measure and manage operational risk using prior knowledge of the causal risk factors and probabilistic reasoning. It is represented in the form of an acyclic graph consisting of nodes and directed arcs. Each node represents a variable that affects or determines operational risk. The arcs denote causal or influential relationships between variables. Each variable in the network is assigned an underlying probability distribution based on subjective prior beliefs. The Bayesian analysis involves improving the prior estimates in the light of additional information about one or more variables in the network.
Application to CA processing
CA processing is one of the most risky and labour-intensive back-office processes in the securities industry. A CA is defined as any action taken by the issuer of a security that affects the structure or financial status of the security.
About 1 million CAs take place every year worldwide over and above the 3 million fixed-rate interest payments and redemptions. Each of these face high operational risks because CA processing is, to a large extent, non-standardised and manual. Failure in handling a single CA has the potential to result in huge losses running into millions of dollars. The global fund management industry alone incurs actual losses of approximately $400 million–900 million every year due to CA processing failures. Indirect losses, arising from incorrect interpretation of CA information, can be significantly larger – such losses cost the securities industry between $2 billion and $9.6 billion annually, according to DTCC (the largest processor of corporate actions and securities clearing, settlement and servicing organisation in the world) and Oxera (one of the largest independent economic consultancies in Europe).
Our article constructs a BBN to model operational loss arising from missed CA announcements (sent by an issuer/vendor). Announcements are typically sent in multiple formats by data vendors and collected by financial organisations for further processing. Since it is largely a manual and non-standardised process, there are chances of missing an announcement, which can result in missed notifications and subsequent settlements, thereby leading to an opportunity/cash loss.
Modelling the network
The first step in the Bayesian process is construction of the operational loss model by identifying the key variables and their cause-effect relationship. The key BBN variables for a missed CA announcement are given below. These variables, in combination, determine operational loss from a missed announcement.
• Data sources. A number of sources including registrar, custodian/sub-custodian and data vendors provide CA announcement data in multiple formats using multiple delivery methods. Further, there is no standard way in which the events are announced by issuers, there is no single securities identification system that is universally accepted and the processing terms and details are often specific to the particular market or financial instrument. Data sources may, therefore, be considered as good or bad.
• CA volumes. Many CAs are announced every year on both equity and debt instruments. Volumes tend to surge during the corporate earnings season, thereby straining the efficiency of the CA staff and increasing the risk of operational loss. The volume of CA announcements to be processed may be low or high.
• CA type complexity. In general, voluntary CAs and mandatory actions with options are considered to be more complex than mandatory actions as they are deadline driven and require processing investor responses. CA type complexity may be in either of the two states – low or high.
• CA processing system. The amount of automation in CA processing varies among organisations. Even the more successful organisations have not managed to automate the entire life cycle. The CA processing system may be assumed to be good or bad.
• Staff efficiency. CA processing continues to be labour-intensive and requires a high degree of manual intervention to resolve exceptions. Staff efficiency may be considered to be in either of the two states – low or high.
The network in figure 1 shows the links between all the BBN variables. The nodes for data sources, CA volumes, CA processing system, CA complexity and staff efficiency represent the causal risk factors or parameters.
The node for 'Loss of missed CA announcement' shows the evidence of operational loss. The magnitude of operational loss due to a missed CA announcement is affected by the data source quality, CA system maturity, staff efficiency and CA volumes.
Assigning probability distributions to BBN variables
The next step in the Bayesian process involves assigning probability distributions to each of the variables in the network. The probability distributions of the variables are provided on the basis of prior knowledge about the behaviour of parameters before operational loss data is observed. (For the sake of simplicity, discrete distributions are considered; however, the technique for assigning continuous distributions is very similar.) In practice, this would involve gathering inputs from the operations staff.
Based on the probability distribution of the causal variables, likelihood estimates for the operational loss are calculated. The results are provided in table C. It should be noted that the probability distribution of 'staff efficiency' is recalculated since it is dependent on two other variables – CA complexity and CA volumes.
Using the above loss estimates, an organisation can estimate the expected loss arising from missing a CA announcement. The expected loss is:
(0.5 × 0.3432) + (1.5 × 0.4349) + (2.5 × 0.2219) = 0.1716 + 0.6524 + 0.5548 = 1.3787 million
Analysis
There are two common techniques of analysing a Bayesian network – scenario analysis and causal analysis.
Scenario analysis involves calibrating one or more causal risk factors in the network and analysing its impact on the loss estimate. For example, an operations manager might be interested in estimating operational losses under heavy processing volumes (all other conditions remaining unchanged). In such a situation, the estimated operational loss is given in table D. It can be seen that, according to the model, processing volumes do not have a significant impact on operational losses. The expected loss given high CA volumes is:
(0.5 × 0.3042) + (1.5 × 0.4606) + (2.5 × 0.2353) = 0.1521 + 0.6909 + 0.5884 = 1.4314 million
Conceptually, causal analysis is the exact opposite of scenario analysis. Under causal analysis, new evidence of operational losses is used to calculate updated probabilities (also referred to as posterior probabilities) of all the causal factors. In other words, additional loss information is propagated to all the nodes in the network. This technique of evidence propagation is extremely useful for analysing the causes of operational losses.
Table E clarifies this concept. If an operations manager is most concerned with large losses ($2 million–3 million) and wants to mitigate this risk, understanding the causes that typically contribute to such losses is important to design a better control infrastructure. From the results in table E, it is evident that the quality of data sources has the maximum impact on operational losses. Consequently, an operations manager, on the basis of these results, might strive to improve the quality of source data – probably by subscribing to more automated feeds.
Refining the network
A BBN may be extended to include decision nodes and utility nodes to facilitate the management of operational risk. A decision node represents a variable controlled by a risk manager to manage operational risk. In the above example, the risk manager may decide to train his staff in order to improve their efficiency. Since the decisions are controlled by the risk manager, they do not have conditional probability tables.
A utility node represents the expected utility from the decision. For instance, the utility node, cost, gives information about the cost associated with training while the utility node, payoff, represents the payoff from increased staff efficiency.
Conclusion
As has been highlighted throughout this article, BBNs provide an effective technique for modelling operational losses. Though they can be used to model almost all operational risk types, they are more appropriate for situations where loss data availability is low. Unlike many statistical techniques, BBNs are investigative in nature – they try to analyse the causes rather than focus solely on the effects. In that respect, it is a forward-looking technique and does not depend entirely on historical losses. This feature makes BBNs particularly effective when the past is not the best predictor of the future.
On the downside, BBNs are somewhat subjective in nature. They are a modeller's view of reality and hence there can be multiple models representing the same operational loss type. Moreover, since business landscapes are dynamic in nature, BBNs involve some amount of maintenance – they need to be regularly updated to incorporate changes in the business. OpRisk
Shilpa Ramamurthy, Harpreet Arora and Anirbid Ghosh work in banking and capital markets solutions consulting at Infosys Technologies. All examples in this paper have been developed using a BBN modelling tool (Hugin Lite) from Hugin Expert A/S. We would like to extend our thanks to Hugin for making this tool freely available for educational purposes
References
Alexander C, 2000–02
Bayesian methods for measuring operational risk
Discussion Paper in Finance
Cruz M, 2002
Modeling, measuring and hedging operational risk
Wiley Finance
Da Costa Lewis N, 2004
Operational risk with Excel and VBA – applied statistical methods for risk management
Wiley Finance
Fenton N, 2002
Incorporating expert judgement in operational risk quantification
Critical Systems Conference
Jensen F, 1996
An introduction to Bayesian networks
Springer-Verlag
Kadie C, D Hovel and E Horvitz, 2001
A component-centric toolkit for modeling and inference with Bayesian networks
Microsoft Research
Yasuda Y, 2003
Application of Bayesian inference to operational risk management
Doctoral theses
1. BBN for a missed CA announcement |
A. Frequency and severity of Basel defined operational risk categories | ||
Risk category | Frequency | Severity |
Internal fraud | Low | High |
External fraud | Low | High |
Employment practices and workplace safety | Low | Low |
Clients, products and business practices | Low/medium | High/medium |
Damage to physical assets | Low | Low/medium |
Business disruption and system failures | Low/medium | Low/medium |
Execution, delivery and process management | High | Low |
B. Prior probability distributions of BBN variables |
Data sources | Probability | CA complexity | Probability | CA volumes | Probability |
Good | 70 | Low | 80 | Low | 30 |
Bad | 30 | High | 20 | High | 70 |
CA processing system | Probability | CA complexity | Low | High | ||
Good | 60 | CA volumes | Low | High | Low | High |
Bad | 40 | Efficiency - Low | 25 | 40 | 40 | 55 |
- High | 75 | 60 | 60 | 45 |
Volumes | Low | |||||||
Processing | Good | Bad | ||||||
Staff efficiency | Low | High | Low | High | ||||
Data sources | Good | Bad | Good | Bad | Good | Bad | Good | Bad |
Loss $0–1 million | 50 | 20 | 60 | 30 | 30 | 10 | 50 | 20 |
1–2 million | 30 | 50 | 30 | 40 | 40 | 50 | 40 | 50 |
2–3 million | 20 | 30 | 10 | 30 | 30 | 40 | 10 | 30 |
Volumes | High | |||||||
Processing | Good | Bad | ||||||
Staff efficiency | Low | High | Low | High | ||||
Data sources | Good | Bad | Good | Bad | Good | Bad | Good | Bad |
Loss $0–1 million | 40 | 10 | 50 | 20 | 20 | 10 | 30 | 10 |
1–2 million | 40 | 60 | 40 | 50 | 50 | 40 | 50 | 50 |
2–3 million | 20 | 30 | 10 | 30 | 30 | 50 | 20 | 40 |
C. Bayesian inference | ||
Staff efficiency | Probability | |
Low | 38.5 | |
High | 61.5 | |
Op loss US$ | Probability | |
0–1 million | 34.32 | |
1–2 million | 43.49 | |
2–3 million | 22.19 |
D.Scenario analysis | ||
CA volumes | Probability | |
Low | 0 | |
High | 100 | |
Op loss US$ | Probability | |
0–1 million | 30.42 | |
1–2 million | 46.06 | |
2–3 million | 23.53 |
E. Causal analysis: maximum operational loss | |||||
Op loss US$ | Probability | Data sources | Probability | CA complexity | Probability |
0–1 million | 0 | Good | 53.58 | Low | 79.02 |
1–2 million | 0 | Bad | 46.42 | High | 20.98 |
2–3 million | 100 |
|
|
|
|
CA volumes | Probability | Staff efficiency | Probability | CA system | Probability |
Low | 25.79 | Low | 48.47 | Good | 50.54 |
High | 74.21 | High | 51.53 | Bad | 49.46 |
Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.
To access these options, along with all other subscription benefits, please contact info@risk.net or view our subscription options here: http://subscriptions.risk.net/subscribe
You are currently unable to print this content. Please contact info@risk.net to find out more.
You are currently unable to copy this content. Please contact info@risk.net to find out more.
Copyright Infopro Digital Limited. All rights reserved.
You may share this content using our article tools. Printing this content is for the sole use of the Authorised User (named subscriber), as outlined in our terms and conditions - https://www.infopro-insight.com/terms-conditions/insight-subscriptions/
If you would like to purchase additional rights please email info@risk.net
Copyright Infopro Digital Limited. All rights reserved.
You may share this content using our article tools. Copying this content is for the sole use of the Authorised User (named subscriber), as outlined in our terms and conditions - https://www.infopro-insight.com/terms-conditions/insight-subscriptions/
If you would like to purchase additional rights please email info@risk.net
More on Operational risk
Financial crime and compliance50 2024
The detailed analysis for the Financial crime and compliance50 considers firms’ technological advances and strategic direction to provide a complete view of how market leaders are driving transformation in this sector
Automating regulatory compliance and reporting
Flaws in the regulation of the banking sector have been addressed initially by Basel III, implemented last year. Financial institutions can comply with capital and liquidity requirements in a natively integrated yet modular environment by utilising…
Investment banks: the future of risk control
This Risk.net survey report explores the current state of risk controls in investment banks, the challenges of effective engagement across the three lines of defence, and the opportunity to develop a more dynamic approach to first-line risk control
Op risk outlook 2022: the legal perspective
Christoph Kurth, partner of the global financial institutions leadership team at Baker McKenzie, discusses the key themes emerging from Risk.net’s Top 10 op risks 2022 survey and how financial firms can better manage and mitigate the impact of…
Emerging trends in op risk
Karen Man, partner and member of the global financial institutions leadership team at Baker McKenzie, discusses emerging op risks in the wake of the Covid‑19 pandemic, a rise in cyber attacks, concerns around conduct and culture, and the complexities of…
Moving targets: the new rules of conduct risk
How are capital markets firms adapting their approaches to monitoring and managing conduct risk following the Covid‑19 pandemic? In a Risk.net webinar in association with NICE Actimize, the panel discusses changing regulatory requirements, the essentials…
Building resilience into ESG risk management
Risk and resilience continue to play an important role in the navigation of an increasingly uncertain world. Fusion Risk Management explores why it is equally crucial for technology to support organisations in addressing pertinent environmental, social…
Operational resilience: charting evolution, strengthening impact
Arming a business in preparation for robust operational resilience measures is not a one-step solution – it continues to evolve. The key to strengthening defences against all events – especially the unlikely but plausible – is to build business agility…
Most read
- Top 10 operational risks for 2024
- Regulators’ FRTB estimates based on faulty premise – industry study
- Top 10 op risks: AI fears drive cyber risk to record high